{ "cells": [ { "cell_type": "markdown", "id": "6aa913a6-3276-42eb-8c31-906a372a9d84", "metadata": {}, "source": [ "# Temporal Cross Conversational Memory for Large Language Models" ] }, { "cell_type": "markdown", "id": "ce93a930-bd64-43be-83ad-a8c9b55b0b70", "metadata": {}, "source": [ "*Sriniketh Ponnada*" ] }, { "cell_type": "markdown", "id": "0cf4b351-0bd9-4336-8cd0-db5d9d70c3de", "metadata": {}, "source": [ "## The problem" ] }, { "cell_type": "markdown", "id": "a8ea23c1-1c6b-400c-a88b-da38728e1029", "metadata": {}, "source": [ "Consumer-facing LLMs hosted by frontier AI labs today offer a feature generally known as \"Personalization\" that allows the LLM to reference information from past conversations and provide results tailored to the user. The visible mechanism used to achieve this is the curation of a summary about the user from past conversations, updated periodically. This summary is either appended directly to the system prompt of every conversation and preloaded into the context window, or individual files on each topic about the user are listed in the system prompt, allowing the model to read each file through a function or tool call. \n", "(For example, see https://claude.ai/new#settings/customize-memory or https://chatgpt.com/#settings/Personalization). \n", "\n", "In this writeup, I will refer to the cross conversational personalization in LLMs as \"memory\". " ] }, { "cell_type": "markdown", "id": "290ee633-c420-4d28-b899-fcce9fdc2b92", "metadata": {}, "source": [ "Memory is an especially important feature for consumer oriented chatbots (in contrast to system facing LLMs used as tools), as the average consumer converses with an LLM as if it were another human, even while understanding otherwise. Therefore, it is helpful to view the behaviour of a chatbot as being analogous to the behaviour of a human in conversation and identifying its shortcomings. \n", "\n", "The first shortcoming is downstream of the concise nature of a summary — chatbots have to tradeoff details for compactness, something that humans do not do. Humans can effortlessly recall niche details from previous conversations, and the user expects this from chatbots as well. A fact that doesn't make it into the summary is invisible the model. Worse — an incorrect fact, or one taken out of context that makes it into memory will be surfaced in future conversations. \n", "\n", "Second — Facts present in a summary are time independent. The model cannot tell old and new information apart. Old information remains in memory even if the information is untrue or expired. Unless the user explicitly notifies that a piece of information is irrelevant, the model will not know to remove old facts from the summary. \n", "\n", "Third — By preloading the summary into every conversation through the system prompt, factually correct but irrelevant facts are introduced to the model, biasing it to include such facts when unnecessary. We do not recollect everything we know about someone every time we converse with them. Recollection is natural and triggered by the details of the conversation itself.\n", "\n", "Fourth — A summary cannot preserve the texture of a memory. 'Bill has a difficult relationship with his boss' loses the specific words and emotions that Bill uses and conveys in the conversation. Therefore, on retrieval, the model does not know how to weigh a fact, or how strongly the user believes in something. It may downplay something sensitive, or bring up minor details the user said in passing. " ] }, { "cell_type": "markdown", "id": "ce797242-8ce5-4e0b-bc53-10304a9ed332", "metadata": {}, "source": [ "## What a better system needs to do" ] }, { "cell_type": "markdown", "id": "83d85bd9-5941-43aa-bfdf-64d8d351ad99", "metadata": {}, "source": [ "Given that users expect chatbot memory to feel human-like, it's reasonable to look at how human recall subjectively behaves as a design inspiration. A fitting analogy to draw on are the parallels between the two types of memory in humans and LLMs. Humans have a short-term memory, allowing us to recall information from the recent past, such as memories from the same day. Short term memory is analogous to the context window of an LLM, within which the model can surface details exactly as they occurred. Similarly, Long term memory in humans — the ability to recollect information from weeks, months or years in the past is analogous to what a good cross conversation memory system in LLMs should look like. \n", "\n", "The key features of Long term memory in humans (from my experience as a human) are: \n", "\n", "1) We as humans can recall past events exactly as they happened.\n", "2) Recollection is a passive task. We recollect information based on the situation we are in, and the triggers that cause it.\n", "3) Recollection is a chain event. We can walk through a chain of related but distinct memories based on their relationships and relevance.\n", "4) Memories are temporal. Old memories fade and decay when they are not recollected, but are strengthened when recalled frequently. The number of times a memory is recalled is in itself a signal for how important it is.\n", "5) It can be hypothesized that the transfer of relevant memories from short term to long term memory happens during sleep. \n", "\n", "A good system for LLM memory should draw on these principles, as they solve the key problems mentioned previously." ] }, { "cell_type": "markdown", "id": "1fcaa94a-f069-41e3-be8b-e10d4bbcd1ea", "metadata": {}, "source": [ "## The Core Idea" ] }, { "cell_type": "markdown", "id": "f013b8c6-0e64-43a8-83a8-c168c9a9f49b", "metadata": {}, "source": [ "Based on the features of human recollection mentioned above, the form of a memory system that incorporates them points towards a knowledge graph. Here's why:\n", "\n", "1) A knowledge graph for this purpose can be built by storing nouns (people, places etc.) as nodes, and verbs as edges, capturing relations between different entities the user comes across. The memory is in the relation, but nodes act as anchor points for easy recollection. It is easier to identify a person on a knowledge graph than to identify a specific memory in a large sea of text.\n", "2) As memories are the relations stored along edges, the importance of a memory can be determined by the weight of the edge. Edges traversed frequently can be made to have high weights, and otherwise, edges can be made to decay with time. \n", "3) If edges between nodes act as memories, then traversal across these edges is recollection. Since related nodes are naturally linked together in a knowledge graph, traversal between related ideas allows chain recollection and captures the context around a specific memory as adjacent edges are traversed.\n", "4) The use of an embedder to automatically trigger a traversal of an edge during a conversation, allows recall to happen automatically, rather than needing the model to call a tool to read the graph.\n", "5) Once built, a knowledge graph can be expanded simply by adding more nodes and edges, rather than requiring a full rewrite every time. The graph can grow indefinitely, without constraints on graph size, because only relevant edges are surfaced during the conversation with the user.\n", "6) By attaching episodic snippets of the exact prompt from which a relation was created to the edge, and loading this into the context window of the LLM during a conversation when the associated edge is traversed, memory recollection is exact, capturing the user in their own words rather than allowing the misinterpretation of a summary." ] }, { "cell_type": "markdown", "id": "57e75414-6037-48c2-a203-cbe37000f8b0", "metadata": {}, "source": [ "## The Architecture" ] }, { "cell_type": "markdown", "id": "02a7eeb5-5e80-4d06-bc9a-3b0ccb07b7e6", "metadata": {}, "source": [ "### How the system works in practice" ] }, { "cell_type": "markdown", "id": "1cd9d996-b5ba-4003-800d-57423c843937", "metadata": {}, "source": [ "At a high level, this is how the memory system works:\n", "\n", "1) The user has their first conversation with the LLM.\n", "2) After the conversation is completed, the text from the full conversation transcript is compiled together and passed to another LLM.\n", "3) This LLM extracts entities from the text (PERSON, ORG, PLACE, EVENT, THING, TOPIC, PREFERENCE, OTHER) and links them via the relations between them. This collection of entities and relations forms a map of the LLM's memory.\n", "4) A snippet of the text from which a relation was extracted is attached to that relation, to provide context.\n", "5) A knowledge graph containing entities as nodes and relations as edges is built out. A self node for the user is designated on first write and used to resolve placeholders.\n", "6) When the user has their next conversation with the model, an embedder scans the user's text at every turn to find words semantically similar to entities and relations on the graph.\n", "7) When a match is found, activation spreads up to 2 hops in both directions from those seed entities, weighted by edge strength, and the strongest neighbourhood is injected into context as relation lines (Entity-Relation-Entity strings) plus the original snippets.\n", "8) The LLM uses this accumulated context to reply to the user.\n", "9) After the conversation completes, the existing graph is appended to and grown further. New nodes and relations are added based on new information from the user. Duplicate nodes are merged and ambiguities resolved. Relations that go unused over time fade in strength, while ones that keep coming up are reinforced.\n", "10) The process repeats." ] }, { "cell_type": "markdown", "id": "1ff2b0ad-8d31-4732-8aa9-8bbe7e0fb43f", "metadata": {}, "source": [ "### Each part" ] }, { "cell_type": "markdown", "id": "7bd533c1-ac6c-4833-bfde-91159f2d487d", "metadata": {}, "source": [ "#### The Schema" ] }, { "cell_type": "markdown", "id": "d71719f6-b921-4b28-b3cf-dba2fa897ad4", "metadata": {}, "source": [ "Memory in this system is a graph, with 3 main parts: \n", "\n", "A **node** is a referent: a person, organization, place, event, thing, topic, preference, or other entity the user has mentioned. There are only those eight kinds, kept fixed so the same idea is always filed in the same way. A node holds a name, alternate names, a small set of stable properties, a confidence score, and a list of which conversations it appeared in. It holds no strength and does not decay. It exists so there is something to attach memories to and search from. \n", "\n", "An **edge** is the memory itself: a single claim linking two nodes, such as Bill → HAS_BOSS → Sarah. Relations are always uppercase, verb first phrases of one to four words. Each edge carries the claim's strength from 0 to 100, the extractor's confidence from 0 to 1, a stability class that sets how fast it should fade (immutable, stable, mutable, time_bound, ephemeral), and a cardinality (one_to_one versus one_to_many) that records whether the relation can truthfully point to more than one target. Only temporary relations can carry an expiry. Nothing is allowed to relate to itself.\n", "\n", "The **snippet** is what makes that claim admissible, and what keeps it honest. Every edge must cite the utterance it came from: the user's own words, stored verbatim on the edge alongside a shorter evidence field. The edge does not describe the snippet; the snippet justifies the edge. Without it, there is no reason for the relation to be there at all. \n", "More importantly, the snippet is what prevents memory from collapsing into a summary. A bare relation like HAS_BOSS says almost nothing about how that fact was expressed. Whether it was said with frustration, with affection, as an aside, or as something dwelled on. The snippet preserves that flavour: the tone, the emphasis, the surrounding context of the sentence it came from. It also carries a useful robustness. The verbatim text inevitably holds details the extractor did not promote into nodes, edges, or properties, but which still matter when the memory is recalled. By holding the snippet, the edge holds on to what structured extraction would otherwise throw away." ] }, { "cell_type": "markdown", "id": "05c5ba5d-f9fe-4335-8345-c8c00f73c267", "metadata": {}, "source": [ "#### Extraction and Graph Construction" ] }, { "cell_type": "markdown", "id": "b6f11330-a5c5-44c2-9d46-4fdd5c649c30", "metadata": {}, "source": [ "Graph construction happens once per conversation, after it finishes. At the end of a conversation, the full transcript is gathered and handed to a language model with one job: propose what should be remembered from it. \n", "The model is asked to return three things in a fixed shape: the entities mentioned, the relations between them, and a short record of the conversation itself including a summary, an importance score, and a few tags. \n", "To help it avoid inventing duplicates, it is also shown the relevant part of the existing graph: the neighbourhood around anything in the current conversation that resembles something already known. That context is found the same way retrieval works, by matching the conversation against what is already stored, except here it is deliberately broad. At write time we would rather show the model too much than too little.\n", "A deterministic pass then cleans what the model returns. Casing is corrected, unknown stability labels fall back to stable, and expiry dates attached to permanent relations are removed. What cannot be repaired is dropped with a note rather than allowed to corrupt the graph. \n", "Relations that point to themselves, relations whose labels are malformed, and relations where neither end refers to anything known do not make it into the graph. When exactly one end is unknown, a provisional node is created for the missing side and marked as low-confidence, so the relation and the concept it points to are preserved rather than lost. \n", "An empty result is valid. Not every conversation contains anything worth keeping. \n", "The outcome is a small, checked batch of new nodes and edges, each edge carrying its verbatim snippet and each node and edge citing the conversation it came from. Nothing is merged or decided here. That batch is passed forward to merge and disambiguation, which decide what is genuinely new and what is a new sighting of something already known." ] }, { "cell_type": "markdown", "id": "e0e780b1-d1cc-4a5f-94cc-0538d1d913b1", "metadata": {}, "source": [ "#### Merge and Disambiguation" ] }, { "cell_type": "markdown", "id": "747992cb-f094-44d1-9e97-d35528fe4307", "metadata": {}, "source": [ "Every new conversation reintroduces people and ideas the system may already know. Extraction proposes names; merge decides whether each one is genuinely new or another sighting of something already on the graph. Merging contains a meaningful asymmetry: merging two different people destroys both. There is no clean way to split them apart later. Missing a merge only creates a duplicate, which can still be fixed. The whole stage is built around that asymmetry: merge quickly when certain, hesitate whenever unsure. \n", "\n", "The first pass is deterministic and needs no model. Each proposed entity is compared only against stored entities of the same kind, so a person named Sam can never collapse into an organization named Sam. Within that set, rules run in order. \n", "1) An exact match on name and kind merges immediately. \n", "2) If the new name appears in an existing node's list of alternate names (aliases), it merges — unless two different nodes both claim it, in which case it is held for review. \n", "3) The same holds for common abbreviations, drawn from a fixed set of groups like William, Bill, and Will: one match merges, several matches mean ambiguity. \n", "4) Weaker resemblances never merge on their own. A spelling similarity above 0.80, (*measured as one minus normalized edit distance*), and a word-overlap test with one name containing the other, or sharing at least half their words each nominate candidates for a second look. If nothing fires at all, the entity is new. \n", "\n", "Generic self-references are the one exception to this caution. When the extractor does not yet know the user's name it emits placeholders like User or me. Since there is structurally only one user per graph, these are forced directly onto the designated self node rather than risking a new placeholder per conversation. That self node itself is chosen once, from the source of the first relation ever observed, on the convention that relations point outward from the user, and then kept fixed. \n", "\n", "Everything held for review goes to a language model with richer context: the proposed entity, the conversation it came from, and each candidate's surrounding neighbourhood on the graph, so shared employers, family, and places count as evidence of identity. \n", "The model returns a confidence per candidate, and a fixed banding decides: above 0.90 merges silently, from 0.70 to 0.90 merges but is flagged for inspection, below 0.70 stays separate. The highest qualifying candidate wins. A soft hint from the extractor, its own guess at which existing node this might be, can only escalate an entity into review, never merge it directly. \n", "\n", "Committing the decision is mechanical. A new node is added with its source conversation attached. A merged node keeps its existing identity and absorbs the new sighting:\n", "\n", "- the new surface form becomes an alias;\n", "- properties are combined;\n", "- the conversation is cited; and\n", "- confidence becomes the higher of the two.\n", "\n", "Edges are then remapped through the same decisions, so a relation pointing at a merged-away name now points at the surviving node. If both ends of a relation land on the same node after remapping, it is dropped. The graph never allows self-relations." ] }, { "cell_type": "markdown", "id": "6a24f8b9-2459-4c8d-95f3-61a8de6c849a", "metadata": {}, "source": [ "#### Decay and Reinforcement" ] }, { "cell_type": "markdown", "id": "4426432a-8233-40e8-ad9a-86061c335d3d", "metadata": {}, "source": [ "Memory here is not stored and left alone. Every relation carries a strength from 0 to 100 that moves over time: it fades when ignored and recovers when used. Nodes never fade. Only relations do, because only relations are memories. \n", "Fading follows a simple exponential curve. The strength after time is the starting strength multiplied by a decay factor that shrinks with each passing day. How fast it shrinks depends on the stability assigned at extraction. Immutable relations do not fade at all. Stable ones fade very slowly, mutable ones faster, and temporary ones fastest of all, with expiry dates allowed only on the last two. A daily pass recomputes every edge from the time since it was last touched. When strength falls below a small threshold, the edge is marked dormant. It is never deleted. A faded memory stays in the graph as provenance, and a later conversation can always revive it. It simply stops competing for recall until then. \n", "\n", "
\n", "In exact form, with $t$ measured in days since the edge was last touched:\n", "\n", "$$ S(t) = S_0 \\cdot e^{-\\lambda t} $$\n", "\n", "where $S_0$ is the strength at last touch (new edges start at $100$),\n", "and $\\lambda$ is set by stability:\n", "\n", "| stability | $\\lambda$ |\n", "|---|---|\n", "| immutable | $0.000$ |\n", "| stable | $0.005$ |\n", "| mutable | $0.020$ |\n", "| time_bound | $0.050$ |\n", "| ephemeral | $0.200$ |\n", "
\n", "\n", "\n", "Reinforcement is the opposite motion and comes from two places. The first is repetition at write time: if a new conversation observes a relation that already exists, the existing edge is strengthened by a fixed amount, capped so it can never exceed its starting value, and its snippet is refreshed only if the new sighting arrived with higher confidence. The second is use at read time. Every read logs which edges its traversal crossed, without changing them. At the close of the conversation those logged edges receive the same fixed bump. In other words, a memory grows stronger both when it is restated and when it proves useful for understanding something else.\n", "The separation matters. Confidence records how sure the system was when it first heard something and never moves on its own. Strength records how alive that thing has stayed since, rising on reuse and falling with neglect. A childhood address can therefore remain high-confidence but long dormant, while a new colleague's name stays weak until it keeps coming up. Time, frequency, and usefulness are all visible in one number, without ever rewriting what was originally said.\n", "\n", "
\n", "Reinforcement, on re-observation or on read-traversal, is:\n", "\n", "$$ S_{\\mathrm{new}} = \\min(S_{\\mathrm{old}} + 15,\\, 100) $$\n", "\n", "An edge with $S < 2.0$ is flagged dormant and never deleted. \n", "
\n", "\n" ] }, { "cell_type": "markdown", "id": "c863d0de-d0a0-4431-af23-17b0b2ff10c1", "metadata": {}, "source": [ "#### Retrieval and Traversal" ] }, { "cell_type": "markdown", "id": "6d11e5a3-265f-4c6d-9c9d-61a77c608bf4", "metadata": {}, "source": [ "Retrieval answers one question per turn: given what the user just said, what small piece of the graph deserves to be in the model's context right now. This is one of the key advantages to using a graph to store memory — the graph can grow indefinitely, but only those parts of the graph that are relevant to the conversation at hand make it into the model's context window. \n", "Retrieval never changes the graph. It only walks it, then logs where it walked so a later write can strengthen those paths. \n", "It starts with words, not vectors. The incoming message is split into content words — lowercase tokens with function words and single letters removed, deduplicated. This is not linguistic analysis; it just concentrates the signal on nouns and other load-bearing words.\n", "Each remaining word is embedded and compared against everything stored. Every node is indexed by its name plus aliases, and every distinct relation label is indexed once as a phrase, so partner can match HAS_PARTNER even when no node is named partner. All vectors are length-normalized, so similarity is a dot product:\n", "$$ \\mathrm{sim}(q, k) = q \\cdot k $$\n", "Because embedding spaces are anisotropic — random pairs already sit well above zero — the comparison is done after subtracting the corpus mean. With $\\mu$ the mean of all stored vectors:\n", "$$ q_c = \\frac{q - \\mu}{\\lVert q - \\mu \\rVert}, \\quad M_c = \\mathrm{normalize}(M - \\mu), \\quad \\mathrm{score} = M_c \\, q_c $$\n", "\n", "This pulls the unrelated baseline back toward zero so a threshold means something. A word seeds on every hit above $0.30$, plus one margin case: a top hit above $0.15$ that beats its runner-up by at least $0.15$ is accepted as a clear winner. Relation hits resolve to the endpoints of every edge carrying that relation. Stale keys are dropped.\n", "Activation then spreads from all seeds at once, up to two hops, in both directions along each edge. A node does not hand its full activation to every neighbour — it divides it in proportion to edge strength, so a heavily connected hub cannot flood the graph:\n", "$$ \\mathrm{contrib}(u \\to v) = \\mathrm{act}(u) \\cdot \\frac{\\mathrm{strength}(u,v)}{\\sum_{e \\ni u} \\mathrm{strength}(e)} \\cdot 0.6 $$\n", "Contributions arriving by different paths add up. Each node's total is remembered along with whether it was reached mostly forward (following edges out), mostly in reverse (following them in), or as a seed, and at which hop. Every edge crossed is logged for deferred reinforcement.\n", "Two adjustments follow:\n", "\n", "1. First, a small co-episode bonus: any node that shares a source conversation with an already-fired node ($\\mathrm{score} \\ge 0.30$) gains $+0.15$, capturing the idea that things mentioned together belong together. Seeds are exempt — they need no bonus.\n", "2. Second, a budget keeps injection terse. Only nodes at or above $0.30$ fire, capped at eight, split roughly $70/30$ between forward-or-seed and reverse, with leftovers backfilling whichever side has room.\n", "Injection itself is tiered. Fired relations between fired nodes are rendered as compact lines — $Bill -[$HAS_BOSS$]\\to Sarah$ — forming the $[MEMORY]$ block. Then, for nodes scoring at or above $0.60$, up to four of their own strongest incident snippets each are added as the $[CONTEXT]$ block, verbatim and deduplicated. Superseded edges are hidden in both. The model therefore sees structure first and texture only where the signal is hot — enough to recall precisely, without re-reading everything known about someone every turn." ] }, { "cell_type": "markdown", "id": "0e20744e-0f2f-4c52-83a5-9c9fd4afca66", "metadata": {}, "source": [ "#### The analogy to synaptic activation" ] }, { "cell_type": "markdown", "id": "ce46fd13-7664-4724-937e-fa9cef33ca9c", "metadata": {}, "source": [ "The resemblance to synapses in the human brain is deliberate but partial. \n", "It holds at the level of dynamics, not mechanism. **What holds** is the use–dependent loop. An edge that fires, crossed during retrieval is later strengthened, and an edge that never fires decays exponentially toward dormancy. That is Hebbian in spirit: what is recalled together stays recallable, what is ignored fades, and frequency of use becomes its own measure of importance. The co-episode bonus is the same idea one step wider: nodes introduced in the same conversation lend each other activation, the graph version of firing together wiring together. Degree normalization plays the role of a limited resource: a hub cannot drive all its neighbours at full strength, so only genuinely convergent paths survive, much as lateral competition constrains biological spread. Even dormancy without deletion has a neural echo — a weakened trace that can be reactivated is closer to depotentiation than to pruning. \n", "\n", "**What does not hold** is almost everything structural. A biological weight is a continuously remodeled conductance shaped by spike timing, calcium, neuromodulation, and sleep-dependent consolidation. Here strength is a single scalar per edge, moved only by two discrete events: a fixed $+15$ bump on reuse, capped at $100$, and a daily exponential multiply set by a hand-chosen stability class. There is no gradient, no timing dependence, no inhibition, and no learning inside retrieval itself. Activation is a transient walk computed fresh each turn, then discarded except for a log of which edges were crossed. The hard cognitive work deciding whether two mentions are the same person, whether a new relation replaces an old one is not done by the network at all but by a language model reasoning over neighbourhoods, with deterministic rules guarding the cases where a mistake would be unrecoverable. Symbols do what weights cannot here: keep identity, tense, and contradiction explicit. \n", "\n", "The graph borrows the economy of synaptic memory: strengthen the useful, let the rest fade, never pay for what is not recalled while keeping meaning itself symbolic and inspectable. Activation behaves like recall; it is not learning. Learning happens only at write time, in daylight, one checked batch per conversation." ] }, { "cell_type": "markdown", "id": "9e6624c4-ef09-45c1-bef6-7f937380015a", "metadata": {}, "source": [ "## An example showcasing the system in action:" ] }, { "cell_type": "markdown", "id": "3f8cdc00-27fc-4d52-9273-b3b8dd2c1c0e", "metadata": {}, "source": [ "Here is a simple 2 turn conversation demonstrating extraction and recall with this system:" ] }, { "cell_type": "markdown", "id": "62d3f888-1ded-460b-8069-40290e14ccdd", "metadata": {}, "source": [ "*Note: This conversation happened on Opencode with GLM 5.3 using the pegasus-opencode npm plugin. Some input text was minimally modified before being pasted here to correct spellings and grammar.*" ] }, { "cell_type": "markdown", "id": "4b433f4a-e58c-4a51-b65a-eb7851f74445", "metadata": {}, "source": [ "### The conversation" ] }, { "cell_type": "markdown", "id": "0b0324f6-40f9-4c90-96d0-0a0cdfb321a2", "metadata": {}, "source": [ "User: \n", "\n", "
\n", "Hey! I'm Trevor, a consultant at Deloitte. I currently work as a financial analyst, but I am conflicted about the next step in my career. I could either transition into a full-time role at one of the tech companies I have been consulting for recently, primarily fast-growing AI startups, or I could switch lanes entirely and join a major investment bank. While I don't know much about AI from a technical standpoint, I deeply understand the financials, and I know it is the next big wave. I also hear that major banks are aggressively trying to position themselves to lead the upcoming influx of AI company IPOs over the next few years. My experience with tech startups could be highly valuable to these large banks, and the cash compensation would be substantial. On the other hand, I could aim for a CFO track at a smaller startup, trading immediate liquidity for significant equity and incredible upside. What do you reckon I should do?
\n", "\n", ". \n", "\n", "Model: \n", "\n", "
\n", "Nice problem to have, Trevor. Here's how I'd think about it:\n", "My lean: the bank, right now. \n", "Your edge — understanding AI startup financials from the inside — is most valuable to banks during the IPO wave, and that edge is temporary. Once the wave matures or cools, everyone will have \"AI experience\" and yours won't be differentiated. Banks paying big cash for exactly your profile is you selling your experience at its peak. Going in-house at a startup, by contrast, puts you in a room full of technical people where finance is a support function, not a value driver. \n", "How to frame it as an analyst (you'll appreciate this): \n", "- Bank = cash-heavy, lower variance, and preserves optionality. Two years of IPO deal experience makes you more valuable to startups later — likely at better terms (CFO of a Series B/C rather than a seed-stage scramble). \n", "- Startup CFO = a concentrated, illiquid position. Your salary, equity, and human capital all sit in one early-stage basket. That's the opposite of diversification, and you of all people know most options expire worthless. \n", "- Full-time at a current client = you'd be joining at a growth-stage valuation, so your equity has already run up; the easy money was made before you got in. \n", "When the other options win: if you have years of runway saved, genuinely believe in one specific company's unit economics (not just the vibe), and can stomach the variance — the CFO path is the only one with life-changing upside. The bank path's ceiling is high cash but no jackpot. \n", "A few questions that would sharpen the answer: How much of a cash cushion do you have? Do you actually know how to evaluate a startup's equity (409a vs. preferred, liquidation preferences)? And can you tolerate banking culture for 2–3 years? If the answer to the last one is no, the whole analysis changes.\n", "
\n", "\n", "User: \n", "\n", "
\n", "Well, I guess the thing I value most is agency. I hate having a boss, and my current boss, Joshua, at Deloitte is the only one I've ever been able to tolerate and work with. I fear that the bureaucracy at a big bank will drive me crazy. I know it sounds counterintuitive; as someone in consulting, the expectation is that I would have already learned to deal with bureaucracies and large corporations, but deep down, I just don't like it.At a smaller company, I am hoping things will be leaner, allowing me to act faster and make big decisions with real skin in the game—something people in my profession are often mocked for lacking. I don't love that stereotype, but I have to admit there is an element of truth to it. Having worked for 15 years, I've built up a few hundred thousand dollars in liquidity and a few hundred thousand more in stocks. I am not in dire need of immediate cash, which makes me wonder if I should just chase liquidity now to start my own venture once I am fully financially independent.The company that offered me the CFO role, Nova AI, already has revenue and is growing at around 24% month-over-month. They are still operating at a loss, which is why they reached out to me for help. The founders are college students, but their revenue is approaching a million dollars a quarter, so the business definitely seems to be taking off. At the bank, however, I would be making over 500,000 dollars a year, which is a massive step up from the 300,000 I earn in my current job. So, what do you think? The immediate money is exciting, but my heart and my brain are completely in conflict.\n", "
\n", "\n", ". \n", "\n", "Model:\n", "\n", "
\n", "Honestly, Trevor, I think your \"heart vs brain\" framing is wrong — and resolving that is the whole answer. \n", "Let me reframe it:\n", "The trade, in numbers. Bank: +200K/year over your current comp. Over 3 years, call it 1-1.5M dollars total. That's what you'd be buying — and you'd be paying for it with the exact thing you say you can't tolerate. You've had exactly one tolerable boss in 15 years; investment banks are the most boss-dense, hierarchy-heavy environments in finance, and you'd be entering at the bottom of a new pecking order as the outsider from consulting. The real risk isn't the bureaucracy — it's that you're miserable in 14 months, your performance slips, and you get pushed out before the money fully materializes. For someone with ~1M in assets, the marginal utility of the next 1M is low. The marginal utility of agency, for you, is clearly very high.\n", "Why Nova fits your actual objective function. You told me your #1 value is agency, and you want to eventually start something of your own. The CFO role is the only option that serves both: real decision rights, skin in the game (fixing the consultant's wound you admitted to), and — most importantly — operating experience, which is the single best prerequisite for founding. Banking teaches you deal process; being CFO teaches you how companies actually run: burn, runway, hiring, pricing, boards. Also note the option asymmetry: Nova CFO → bank later is very plausible (an AI-company operator stays hot through the IPO wave); bank → founder later means 3 more years of not building the muscle. And 24% MoM at $1M/quarter is a ~10x annual trajectory if it held — even decaying, that's the kind of curve that makes CFO equity meaningful, and it puts you inside the AI wave you're excited about rather than selling it from the other side of the table.\n", "But here's the catch — Nova's agency isn't guaranteed, it's dilutable. You'd still have bosses: two college students and, eventually, a board. The risk isn't the title, it's whether you have real authority. Before you say yes, diligence it like a deal:\n", "- Cap table: last round priced? Liquidation preferences stacked on top of your common? How much dilution is coming?\n", "- Your deal: what % equity, what vesting, what severance if the founders and you clash.\n", "- Your authority: do you sign off on spend, do you attend board meetings, does the CEO actually defer to you on finance?\n", "- The business: runway in months, gross margin, churn/net revenue retention, is the 24% growth paid or organic (if it's bought with unprofitable acquisition, that's a treadmill, not traction).\n", "- The founders: can they handle pressure? Have they worked with senior operators before, or will they resent the adult in the room? Ask their existing investors this directly — investors will be surprisingly candid about founder maturity.\n", "Take the bank only if diligence fails (they won't share the cap table, runway is under 12 months, growth is paid, the founders bristle at your questions), or if you do the honest self-check and conclude that what you actually want is wealth maximization, not agency — those are both legitimate, but they're different answers.\n", "One last thing, gently: \"I'll go after liquidity and start something when I'm fully independent\" — that horizon recedes forever. You already have enough to survive a failure. People who wait until they feel rich enough to take a risk almost never take it. If Nova's diligence checks out, this is the moment you were saving for.\n", "
" ] }, { "cell_type": "markdown", "id": "29b4000b-1af3-432b-91c5-3538e5c8b7b6", "metadata": {}, "source": [ "### The knowledge graph produced by our system from this conversation" ] }, { "cell_type": "code", "execution_count": 6, "id": "02742544-3e1d-4a49-92d8-170e70622027", "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "\u001b[33mWARNING: Ignoring invalid distribution -umpy (/Users/stponnada/Library/Python/3.9/lib/python/site-packages)\u001b[0m\u001b[33m\n", "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution -umpy (/Users/stponnada/Library/Python/3.9/lib/python/site-packages)\u001b[0m\u001b[33m\n", "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution -umpy (/Users/stponnada/Library/Python/3.9/lib/python/site-packages)\u001b[0m\u001b[33m\n", "\u001b[0m\u001b[33mWARNING: Ignoring invalid distribution -umpy (/Users/stponnada/Library/Python/3.9/lib/python/site-packages)\u001b[0m\u001b[33m\n", "\u001b[0m\n", "\u001b[1m[\u001b[0m\u001b[34;49mnotice\u001b[0m\u001b[1;39;49m]\u001b[0m\u001b[39;49m A new release of pip is available: \u001b[0m\u001b[31;49m25.2\u001b[0m\u001b[39;49m -> \u001b[0m\u001b[32;49m26.0.1\u001b[0m\n", "\u001b[1m[\u001b[0m\u001b[34;49mnotice\u001b[0m\u001b[1;39;49m]\u001b[0m\u001b[39;49m To update, run: \u001b[0m\u001b[32;49m/Library/Developer/CommandLineTools/usr/bin/python3 -m pip install --upgrade pip\u001b[0m\n", "Note: you may need to restart the kernel to use updated packages.\n", "16 nodes, 21 edges\n" ] }, { "data": { "text/html": [ "" ], "text/plain": [ "" ] }, "execution_count": 6, "metadata": {}, "output_type": "execute_result" } ], "source": [ "%pip install -q pyvis\n", "\n", "import json\n", "from IPython.display import HTML\n", "from pyvis.network import Network\n", "\n", "# Embedded graph data (the exact result of the two-turn Trevor/Deloitte/\n", "# Joshua/Nova AI conversation above: 16 nodes, 21 edges) -- no external\n", "# file needed, this cell is fully self-contained.\n", "GRAPH_JSON = '{\"nodes\":[{\"id\":\"PERSON::trevor\",\"label\":\"Trevor\",\"kind\":\"PERSON\",\"aliases\":[],\"properties\":{\"liquid_assets\":\"a few hundred thousand USD (as of 2026-09)\",\"profession\":\"consultant and financial analyst\",\"stock_holdings\":\"a few hundred thousand USD, on top of liquidity (as of 2026-09)\",\"years_of_experience\":\"15 years (as of 2026-09)\"},\"confidence\":1.0},{\"id\":\"ORG::deloitte\",\"label\":\"Deloitte\",\"kind\":\"ORG\",\"aliases\":[\"Deloitte Consulting\"],\"properties\":{\"industry\":\"professional services / management consulting\"},\"confidence\":0.95},{\"id\":\"PERSON::joshua\",\"label\":\"Joshua\",\"kind\":\"PERSON\",\"aliases\":[\"Joshua (Trevor\\'s boss at Deloitte)\"],\"properties\":{},\"confidence\":0.9},{\"id\":\"PREFERENCE::agency_workplace_autonomy\",\"label\":\"Agency (workplace autonomy)\",\"kind\":\"PREFERENCE\",\"aliases\":[\"agency\",\"autonomy\",\"workplace autonomy\"],\"properties\":{\"description\":\"Trevor\\'s top stated value: autonomy over his work and decisions\"},\"confidence\":0.95},{\"id\":\"PREFERENCE::working_without_a_boss\",\"label\":\"Working without a boss\",\"kind\":\"PREFERENCE\",\"aliases\":[\"no boss\",\"being boss-free\",\"avoiding bosses\"],\"properties\":{\"intensity\":\"strong \\\\u2014 stated as \\\\\"I hate having a boss\\\\\"\"},\"confidence\":0.9},{\"id\":\"PREFERENCE::low_bureaucracy_work_environments\",\"label\":\"Low-bureaucracy work environments\",\"kind\":\"PREFERENCE\",\"aliases\":[\"lean environments\",\"minimal bureaucracy\",\"low-bureaucracy environments\"],\"properties\":{\"scope\":\"large companies and banks; never acclimated to them despite a consulting career\"},\"confidence\":0.85},{\"id\":\"PREFERENCE::skin_in_the_game_decision_making\",\"label\":\"Skin in the game (decision-making)\",\"kind\":\"PREFERENCE\",\"aliases\":[\"skin in the game\",\"ownership of big decisions\",\"decision ownership\"],\"properties\":{\"description\":\"making big decisions with personal ownership and real stakes\"},\"confidence\":0.85},{\"id\":\"TOPIC::artificial_intelligence\",\"label\":\"Artificial intelligence\",\"kind\":\"TOPIC\",\"aliases\":[\"AI\"],\"properties\":{},\"confidence\":0.9},{\"id\":\"THING::ai_startups_trevor_s_consulting_clients\",\"label\":\"AI startups (Trevor\\'s consulting clients)\",\"kind\":\"THING\",\"aliases\":[\"tech startups Trevor consults at\",\"fast-growing AI startups (Trevor\\'s clients)\"],\"properties\":{\"profile\":\"mainly AI startups growing really fast; Trevor\\'s recent consulting engagements as of 2026-09\"},\"confidence\":0.9},{\"id\":\"ORG::nova_ai\",\"label\":\"Nova AI\",\"kind\":\"ORG\",\"aliases\":[\"NovaAI\"],\"properties\":{\"founders\":\"college students, who reached out to Trevor for help\",\"growth_rate\":\"around 24% month-over-month (as of 2026-09)\",\"profitability\":\"still loss-making (as of 2026-09)\",\"quarterly_revenue\":\"almost $1M per quarter (as of 2026-09)\",\"stage\":\"early-stage, revenue-generating startup\"},\"confidence\":0.95},{\"id\":\"THING::nova_ai_cfo_offer\",\"label\":\"Nova AI CFO offer\",\"kind\":\"THING\",\"aliases\":[\"CFO offer at Nova AI\",\"Nova AI CFO role\"],\"properties\":{\"status\":\"extended to Trevor; under active consideration as of Sep 2026\",\"title\":\"Chief Financial Officer (CFO)\"},\"confidence\":0.95},{\"id\":\"THING::large_bank_career_path\",\"label\":\"Large-bank career path\",\"kind\":\"THING\",\"aliases\":[\"big bank job\",\"large banking firm path\",\"bank path\"],\"properties\":{\"description\":\"option to leave consulting for a full-time role at a large banking firm\"},\"confidence\":0.85},{\"id\":\"TOPIC::ai_company_ipo_wave\",\"label\":\"AI company IPO wave\",\"kind\":\"TOPIC\",\"aliases\":[\"AI IPO wave\",\"wave of AI company IPOs\"],\"properties\":{\"description\":\"anticipated wave of AI-company IPOs that banks are competing to organize, expected within months to years (per Trevor, Sep 2026)\"},\"confidence\":0.7},{\"id\":\"PREFERENCE::founding_his_own_company\",\"label\":\"Founding his own company\",\"kind\":\"PREFERENCE\",\"aliases\":[\"starting something on his own\",\"own venture\",\"entrepreneurship\"],\"properties\":{\"tentativeness\":\"hedged with \\\\\"maybe\\\\\"; something he wonders about (Sep 2026)\",\"timing\":\"when fully independent financially\"},\"confidence\":0.6},{\"id\":\"TOPIC::financial_side_of_ai\",\"label\":\"Financial side of AI\",\"kind\":\"TOPIC\",\"aliases\":[\"AI financials\",\"financial side of AI companies\"],\"properties\":{\"description\":\"the financial/business side of AI companies, as distinct from the technical side\"},\"confidence\":0.85},{\"id\":\"THING::full_time_ai_startup_career_path\",\"label\":\"Full-time AI startup career path\",\"kind\":\"THING\",\"aliases\":[\"AI startup full-time path\",\"joining an AI startup full-time\"],\"properties\":{\"description\":\"career option of taking a full-time job at the tech companies Trevor consults for, mainly fast-growing AI startups\"},\"confidence\":0.75}],\"edges\":[{\"id\":\"PERSON::trevor::WORKS_AT::ORG::deloitte\",\"from\":\"PERSON::trevor\",\"to\":\"ORG::deloitte\",\"relation\":\"WORKS_AT\",\"strength\":99.95,\"confidence\":0.95,\"stability\":\"mutable\",\"cardinality\":\"one_to_one\",\"snippet\":\"I\\'m Trevor, a consultant at Deloitte. I work as a financial analyst right now, and I am confused as to what I should do next in my career.\",\"properties\":{\"compensation\":\"$300,000/year (as of 2026-09)\",\"role\":\"consultant\",\"title\":\"financial analyst\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:11:04.175643+00:00\"},{\"id\":\"PERSON::trevor::REPORTS_TO::PERSON::joshua\",\"from\":\"PERSON::trevor\",\"to\":\"PERSON::joshua\",\"relation\":\"REPORTS_TO\",\"strength\":99.95,\"confidence\":0.9,\"stability\":\"mutable\",\"cardinality\":\"one_to_one\",\"snippet\":\"I hate having a boss, and my currect boss Joshua at Deloitte is the only one Ive ever been able to tolerate and work with.\",\"properties\":{\"notability\":\"the only boss Trevor has ever been able to tolerate and work with (across a 15-year career)\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:11:09.067083+00:00\"},{\"id\":\"PERSON::joshua::WORKS_AT::ORG::deloitte\",\"from\":\"PERSON::joshua\",\"to\":\"ORG::deloitte\",\"relation\":\"WORKS_AT\",\"strength\":99.95,\"confidence\":0.85,\"stability\":\"mutable\",\"cardinality\":\"one_to_one\",\"snippet\":\"my currect boss Joshua at Deloitte is the only one Ive ever been able to tolerate and work with.\",\"properties\":{\"role\":\"manages Trevor; Trevor\\'s current boss\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:11:11.424810+00:00\"},{\"id\":\"PERSON::trevor::VALUES_MOST::PREFERENCE::agency_workplace_autonomy\",\"from\":\"PERSON::trevor\",\"to\":\"PREFERENCE::agency_workplace_autonomy\",\"relation\":\"VALUES_MOST\",\"strength\":99.99,\"confidence\":0.95,\"stability\":\"stable\",\"cardinality\":\"one_to_one\",\"snippet\":\"Well, I guess the thing I value the most is agency. I hate having a boss, and my currect boss Joshua at Deloitte is the only one Ive ever been able to tolerate and work with.\",\"properties\":{\"rank\":\"single most valued thing (stated Sep 2026)\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:11:16.623347+00:00\"},{\"id\":\"PERSON::trevor::PREFERS::PREFERENCE::working_without_a_boss\",\"from\":\"PERSON::trevor\",\"to\":\"PREFERENCE::working_without_a_boss\",\"relation\":\"PREFERS\",\"strength\":99.99,\"confidence\":0.9,\"stability\":\"stable\",\"cardinality\":\"one_to_many\",\"snippet\":\"I hate having a boss, and my currect boss Joshua at Deloitte is the only one Ive ever been able to tolerate and work with.\",\"properties\":{\"intensity\":\"hates having a boss; in 15 years has tolerated only one (Joshua)\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:11:22.861831+00:00\"},{\"id\":\"PERSON::trevor::PREFERS::PREFERENCE::low_bureaucracy_work_environments\",\"from\":\"PERSON::trevor\",\"to\":\"PREFERENCE::low_bureaucracy_work_environments\",\"relation\":\"PREFERS\",\"strength\":99.99,\"confidence\":0.85,\"stability\":\"stable\",\"cardinality\":\"one_to_many\",\"snippet\":\"I fear that the bureaucracy at a big bank will drive me crazy. I know it sounds counterintuitive. As someone in consulting, I guess the expectation is that I would have already learnt to deal with bureaucracies and large companies, but deep down, I dont like it.\",\"properties\":{\"context\":\"counterintuitive for a consultant, who is expected to have learned to deal with bureaucracies and large companies; deep down he dislikes them\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:11:33.303842+00:00\"},{\"id\":\"PERSON::trevor::SEEKS::PREFERENCE::skin_in_the_game_decision_making\",\"from\":\"PERSON::trevor\",\"to\":\"PREFERENCE::skin_in_the_game_decision_making\",\"relation\":\"SEEKS\",\"strength\":99.99,\"confidence\":0.8,\"stability\":\"stable\",\"cardinality\":\"one_to_many\",\"snippet\":\"At a smaller company, I was hoping things would be leaner and I would be able to act faster and make big decisions with skin in the game, something people in my profession are mocked for not having. I don\\'t like that framing, but I have to admit than an element of it s true.\",\"properties\":{\"context\":\"people in his profession (consulting) are mocked for lacking skin in the game; he dislikes that framing but admits an element of it is true\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:11:41.155277+00:00\"},{\"id\":\"PERSON::trevor::LACKS_TECHNICAL_KNOWLEDGE_OF::TOPIC::artificial_intelligence\",\"from\":\"PERSON::trevor\",\"to\":\"TOPIC::artificial_intelligence\",\"relation\":\"LACKS_TECHNICAL_KNOWLEDGE_OF\",\"strength\":99.95,\"confidence\":0.9,\"stability\":\"mutable\",\"cardinality\":\"one_to_many\",\"snippet\":\"I don\\'t know much about AI technically, but I understand the financials, and I know it\\'s the next big thing.\",\"properties\":{\"level\":\"not much technical knowledge (self-assessed, Sep 2026)\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:11:48.874371+00:00\"},{\"id\":\"PERSON::trevor::CONSULTS_FOR::THING::ai_startups_trevor_s_consulting_clients\",\"from\":\"PERSON::trevor\",\"to\":\"THING::ai_startups_trevor_s_consulting_clients\",\"relation\":\"CONSULTS_FOR\",\"strength\":99.95,\"confidence\":0.85,\"stability\":\"mutable\",\"cardinality\":\"one_to_many\",\"snippet\":\"I can either get a full time job at the tech companies I am consulting at recently (mainly AI startups that are growing really fast), or I could switch lanes and work for a large banking firm.\",\"properties\":{\"client_profile\":\"mainly AI startups that are growing really fast\",\"recency\":\"recent engagements as of 2026-09\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:12:14.594998+00:00\"},{\"id\":\"THING::nova_ai_cfo_offer::OFFERED_BY::ORG::nova_ai\",\"from\":\"THING::nova_ai_cfo_offer\",\"to\":\"ORG::nova_ai\",\"relation\":\"OFFERED_BY\",\"strength\":100.0,\"confidence\":0.95,\"stability\":\"immutable\",\"cardinality\":\"one_to_one\",\"snippet\":\"The company where I have been offered a role as CFO (Nova AI) has revenue already, and they are growing at around 24 percent month over month, but they are still in a loss, and have reached out to me for help.\",\"properties\":{\"origin\":\"Nova AI\\'s founders (college students) reached out to Trevor for help\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:14:12.462727+00:00\"},{\"id\":\"PERSON::trevor::CONSIDERING::THING::nova_ai_cfo_offer\",\"from\":\"PERSON::trevor\",\"to\":\"THING::nova_ai_cfo_offer\",\"relation\":\"CONSIDERING\",\"strength\":99.89,\"confidence\":0.9,\"stability\":\"time_bound\",\"cardinality\":\"one_to_many\",\"snippet\":\"Otherwise, I could become a CFO at a smaller company and trade liquidity for a lot of stock and increddible upside. ... The company where I have been offered a role as CFO (Nova AI) has revenue already, and they are growing at around 24 percent month over month, but they are still in a loss, and have reached out to me for help.\",\"properties\":{\"hoped_for\":\"leaner environment, faster action, big decisions with skin in the game\",\"status\":\"actively weighing vs the bank path; \\\\\"my heart and brain are in conflict\\\\\"\",\"tradeoff\":\"trades liquidity/salary for stock and upside\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:15:48.930343+00:00\"},{\"id\":\"PERSON::trevor::CONSIDERING::THING::large_bank_career_path\",\"from\":\"PERSON::trevor\",\"to\":\"THING::large_bank_career_path\",\"relation\":\"CONSIDERING\",\"strength\":99.9,\"confidence\":0.85,\"stability\":\"time_bound\",\"cardinality\":\"one_to_many\",\"snippet\":\"At the bank though, I will be making over $500,000 a year, which is a big step up from the $300,000 Ive been earning in my current job. So what do you think? The immedate money seems exciting, but my heart and brain are in conflict\",\"properties\":{\"compensation\":\"over $500,000/year, largely cash (\\\\\"I could get paid a lot in cash\\\\\")\",\"rationale\":\"his AI-startup consulting experience would be valuable to big banks\",\"tension\":\"the immediate money seems exciting, but his heart and brain are in conflict\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:16:25.348819+00:00\"},{\"id\":\"THING::large_bank_career_path::CONFLICTS_WITH::PREFERENCE::low_bureaucracy_work_environments\",\"from\":\"THING::large_bank_career_path\",\"to\":\"PREFERENCE::low_bureaucracy_work_environments\",\"relation\":\"CONFLICTS_WITH\",\"strength\":99.9,\"confidence\":0.85,\"stability\":\"time_bound\",\"cardinality\":\"one_to_many\",\"snippet\":\"I fear that the bureaucracy at a big bank will drive me crazy. I know it sounds counterintuitive. As someone in consulting, I guess the expectation is that I would have already learnt to deal with bureaucracies and large companies, but deep down, I dont like it.\",\"properties\":{\"fear\":\"Trevor fears the bureaucracy at a big bank would drive him crazy\",\"note\":\"counterintuitive for a consultant expected to be used to large-company bureaucracy; deep down he dislikes it\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:16:32.226976+00:00\"},{\"id\":\"PERSON::trevor::ANTICIPATES::TOPIC::ai_company_ipo_wave\",\"from\":\"PERSON::trevor\",\"to\":\"TOPIC::ai_company_ipo_wave\",\"relation\":\"ANTICIPATES\",\"strength\":99.9,\"confidence\":0.65,\"stability\":\"time_bound\",\"cardinality\":\"one_to_many\",\"snippet\":\"I know it\\'s the next big thing, and I hear that many banks are trying to be the ones that organize the big wave of AI company IPOs in the next few months and years.\",\"properties\":{\"source\":\"hearsay (\\\\\"I hear\\\\\"), as of Sep 2026\",\"timing\":\"next few months and years\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:16:35.025781+00:00\"},{\"id\":\"THING::large_bank_career_path::MOTIVATED_BY::TOPIC::ai_company_ipo_wave\",\"from\":\"THING::large_bank_career_path\",\"to\":\"TOPIC::ai_company_ipo_wave\",\"relation\":\"MOTIVATED_BY\",\"strength\":99.9,\"confidence\":0.7,\"stability\":\"time_bound\",\"cardinality\":\"one_to_many\",\"snippet\":\"I hear that many banks are trying to be the ones that organize the big wave of AI company IPOs in the next few months and years. My experieence at teech startups might be valuable for these big banks and I could get paid a lot in cash.\",\"properties\":{\"rationale\":\"banks are competing to organize the coming AI-company IPO wave, so Trevor\\'s AI-startup consulting experience would be valuable to them; plus high cash pay\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:16:38.744980+00:00\"},{\"id\":\"PERSON::trevor::ASPIRES_TO::PREFERENCE::founding_his_own_company\",\"from\":\"PERSON::trevor\",\"to\":\"PREFERENCE::founding_his_own_company\",\"relation\":\"ASPIRES_TO\",\"strength\":99.96,\"confidence\":0.6,\"stability\":\"mutable\",\"cardinality\":\"one_to_many\",\"snippet\":\"I am not in the dire need for money, but I wonder if I should just go after liquidity and maybe start something on my own when I am fully independent.\",\"properties\":{\"status\":\"tentative \\\\u2014 hedged with \\\\\"maybe\\\\\"; something he wonders about\",\"timing\":\"once fully independent financially\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:16:41.438474+00:00\"},{\"id\":\"THING::nova_ai_cfo_offer::ALIGNS_WITH::PREFERENCE::skin_in_the_game_decision_making\",\"from\":\"THING::nova_ai_cfo_offer\",\"to\":\"PREFERENCE::skin_in_the_game_decision_making\",\"relation\":\"ALIGNS_WITH\",\"strength\":99.9,\"confidence\":0.7,\"stability\":\"time_bound\",\"cardinality\":\"one_to_many\",\"snippet\":\"At a smaller company, I was hoping things would be leaner and I would be able to act faster and make big decisions with skin in the game, something people in my profession are mocked for not having.\",\"properties\":{\"basis\":\"Trevor\\'s stated hope about smaller companies, applied to the Nova AI CFO option; not a confirmed attribute of Nova AI\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:16:44.103175+00:00\"},{\"id\":\"PERSON::trevor::UNDERSTANDS::TOPIC::financial_side_of_ai\",\"from\":\"PERSON::trevor\",\"to\":\"TOPIC::financial_side_of_ai\",\"relation\":\"UNDERSTANDS\",\"strength\":99.99,\"confidence\":0.85,\"stability\":\"stable\",\"cardinality\":\"one_to_many\",\"snippet\":\"I don\\'t know much about AI technically, but I understand the financials, and I know it\\'s the next big thing.\",\"properties\":{\"belief\":\"considers AI the next big thing\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:16:53.312450+00:00\"},{\"id\":\"TOPIC::financial_side_of_ai::ASPECT_OF::TOPIC::artificial_intelligence\",\"from\":\"TOPIC::financial_side_of_ai\",\"to\":\"TOPIC::artificial_intelligence\",\"relation\":\"ASPECT_OF\",\"strength\":100.0,\"confidence\":0.9,\"stability\":\"immutable\",\"cardinality\":\"one_to_one\",\"snippet\":\"I don\\'t know much about AI technically, but I understand the financials, and I know it\\'s the next big thing.\",\"properties\":{},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:16:55.672083+00:00\"},{\"id\":\"PERSON::trevor::CONSIDERING::THING::full_time_ai_startup_career_path\",\"from\":\"PERSON::trevor\",\"to\":\"THING::full_time_ai_startup_career_path\",\"relation\":\"CONSIDERING\",\"strength\":99.9,\"confidence\":0.75,\"stability\":\"time_bound\",\"cardinality\":\"one_to_many\",\"snippet\":\"I can either get a full time job at the tech companies I am consulting at recently (mainly AI startups that are growing really fast), or I could switch lanes and work for a large banking firm.\",\"properties\":{\"status\":\"one of three options under consideration, Sep 2026; the vaguest of the three\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:17:49.672560+00:00\"},{\"id\":\"THING::full_time_ai_startup_career_path::WOULD_JOIN::THING::ai_startups_trevor_s_consulting_clients\",\"from\":\"THING::full_time_ai_startup_career_path\",\"to\":\"THING::ai_startups_trevor_s_consulting_clients\",\"relation\":\"WOULD_JOIN\",\"strength\":99.9,\"confidence\":0.8,\"stability\":\"time_bound\",\"cardinality\":\"one_to_many\",\"snippet\":\"I can either get a full time job at the tech companies I am consulting at recently (mainly AI startups that are growing really fast), or I could switch lanes and work for a large banking firm.\",\"properties\":{\"note\":\"the option is to convert his current consulting engagements into full-time employment\"},\"dormant\":false,\"superseded_by\":null,\"updated_at\":\"2026-09-14T18:17:52.348746+00:00\"}]}\\n'\n", "graph = json.loads(GRAPH_JSON)\n", "print(f\"{len(graph['nodes'])} nodes, {len(graph['edges'])} edges\")\n", "\n", "# Same palette as engine/src/ontomem/static/viewer.html's KIND_COLOR (light\n", "# mode values), so this matches the live in-app viewer's colors.\n", "KIND_COLOR = {\n", " \"PERSON\": \"#2a78d6\",\n", " \"ORG\": \"#eb6834\",\n", " \"PLACE\": \"#1baf7a\",\n", " \"EVENT\": \"#eda100\",\n", " \"TOPIC\": \"#e87ba4\",\n", " \"PREFERENCE\": \"#008300\",\n", " \"THING\": \"#4a3aa7\",\n", " \"OTHER\": \"#e34948\",\n", "}\n", "\n", "\n", "def node_tooltip(node):\n", " lines = [f\"{node['label']} ({node['kind']})\"]\n", " if node.get(\"aliases\"):\n", " lines.append(\"aliases: \" + \", \".join(node[\"aliases\"]))\n", " for k, v in (node.get(\"properties\") or {}).items():\n", " lines.append(f\"{k}: {v}\")\n", " lines.append(f\"confidence: {node['confidence']}\")\n", " return \"
\".join(lines)\n", "\n", "\n", "def edge_tooltip(edge):\n", " lines = [\n", " f\"{edge['relation']}\",\n", " f\"strength: {edge['strength']} \\u00b7 confidence: {edge['confidence']}\",\n", " f\"stability: {edge['stability']} \\u00b7 cardinality: {edge['cardinality']}\",\n", " ]\n", " for k, v in (edge.get(\"properties\") or {}).items():\n", " lines.append(f\"{k}: {v}\")\n", " if edge.get(\"snippet\"):\n", " lines.append(f\"\\u201c{edge['snippet']}\\u201d\")\n", " return \"
\".join(lines)\n", "\n", "\n", "net = Network(\n", " height=\"750px\",\n", " width=\"100%\",\n", " directed=True,\n", " notebook=False,\n", " cdn_resources=\"in_line\", # embed vis.js itself -- no external CDN dependency\n", " bgcolor=\"#F5F5F5\",\n", " font_color=\"#000003\",\n", ")\n", "\n", "for node in graph[\"nodes\"]:\n", " net.add_node(\n", " node[\"id\"],\n", " label=node[\"label\"],\n", " title=node_tooltip(node),\n", " color=KIND_COLOR.get(node[\"kind\"], KIND_COLOR[\"OTHER\"]),\n", " shape=\"dot\",\n", " size=18,\n", " # Extra fields beyond what vis-network needs for rendering -- pyvis\n", " # passes any unrecognized kwarg straight through onto the node\n", " # object, so the click handler injected below can read the full\n", " # structured record back out (same data the tooltip above is built\n", " # from), not just re-parse the tooltip's HTML.\n", " kind=node[\"kind\"],\n", " aliasesText=\", \".join(node.get(\"aliases\") or []),\n", " props=node.get(\"properties\") or {},\n", " confRaw=node[\"confidence\"],\n", " )\n", "\n", "for edge in graph[\"edges\"]:\n", " net.add_edge(\n", " edge[\"from\"],\n", " edge[\"to\"],\n", " label=edge[\"relation\"],\n", " title=edge_tooltip(edge),\n", " color=\"#8a8a8a\" if not edge.get(\"dormant\") else \"#3a3a3a\",\n", " arrows=\"to\",\n", " width=1 + edge[\"strength\"] / 50,\n", " font={\"size\": 10, \"color\": \"#b0b0b0\", \"strokeWidth\": 0},\n", " relation=edge[\"relation\"],\n", " strengthRaw=edge[\"strength\"],\n", " confidence=edge[\"confidence\"],\n", " stability=edge[\"stability\"],\n", " cardinality=edge[\"cardinality\"],\n", " props=edge.get(\"properties\") or {},\n", " snippetText=edge.get(\"snippet\") or \"\",\n", " )\n", "\n", "net.set_options(\"\"\"\n", "{\n", " \"physics\": {\n", " \"barnesHut\": {\n", " \"gravitationalConstant\": -9000,\n", " \"springLength\": 150,\n", " \"springConstant\": 0.04,\n", " \"avoidOverlap\": 0.4\n", " },\n", " \"stabilization\": { \"iterations\": 300 }\n", " },\n", " \"interaction\": { \"hover\": true, \"tooltipDelay\": 120 }\n", "}\n", "\"\"\")\n", "\n", "# Sidebar/\"Graph Inspector\" behavior -- mirrors engine/src/ontomem/static/\n", "# viewer.html's own inspector drawer (same field layout: kind/name/\n", "# confidence/aliases/properties for nodes, relation/from/to/strength/\n", "# confidence/stability/cardinality/properties/snippet for edges). Injected\n", "# as raw CSS+JS right before in the generated HTML -- `nodes`,\n", "# `edges`, and `network` are pyvis's own top-level `var`s in that same\n", "# document, so this script reuses them directly rather than redeclaring\n", "# its own copies of the graph data.\n", "SIDEBAR_HTML = \"\"\"\n", "
\n", "

Graph Inspector

\n", "
Click a node or edge to see its details.
\n", "
\n", "\n", "\n", "\"\"\"\n", "\n", "graph_html = net.generate_html()\n", "graph_html = graph_html.replace(\"\", SIDEBAR_HTML + \"\")\n", "\n", "escaped = (\n", " graph_html.replace(\"&\", \"&\").replace('\"', \""\").replace(\"<\", \"<\").replace(\">\", \">\")\n", ")\n", "HTML(f'')" ] }, { "cell_type": "markdown", "id": "0fc252ee-3825-4125-a391-0933c5ec9715", "metadata": {}, "source": [ "### Testing Recall" ] }, { "cell_type": "markdown", "id": "7546286e-bf03-4a10-84ea-9300ef07c3a2", "metadata": {}, "source": [ "Here is the fun part. With a system that captures every data point, you can start conversations cold, without any context, and the model will stick pick up right where it left off:" ] }, { "cell_type": "markdown", "id": "2c6eb8a2-07b4-4c14-81b8-91bd6bc53f48", "metadata": {}, "source": [ "(This happened in a fresh conversation. We bring up the manager by name, the specifics of the role, and a personal preference that was captured in the snippets)" ] }, { "cell_type": "markdown", "id": "9ef572dd-d01b-418d-9404-f6d8aa8c93d5", "metadata": {}, "source": [ "User:
I just spoke to joshua. He did encourage me to take the CFO role instead of taking the job at a big bank which I clearly didnt like!
" ] }, { "cell_type": "markdown", "id": "12075e93-1d22-467e-b6f2-e5925d06fd80", "metadata": {}, "source": [ "Model: \n", "
\n", "That's a meaningful vote of confidence — especially coming from Joshua, the only boss you've ever actually worked well with. His logic tracks too: the bank money is tempting, but you'd be trading one tolerable boss for an unknown one, while Nova AI offers you equity-level upside, a CFO title, and no one above you but two college founders who came to you for help.\n", "The real question now is probably the loss-making financials — do you want to dig into Nova AI's numbers (burn rate, runway, margin structure) before you commit?\n", "
" ] }, { "cell_type": "markdown", "id": "beaf5ae7-9110-4b52-9f4d-05afbbd4c7ae", "metadata": {}, "source": [ "(The model picked off right where it ended, and even brought in specifics such as its concerns about the startup's burn rate and other economics)" ] }, { "cell_type": "markdown", "id": "dd8cc82b-ffcc-4027-89f7-330783ac3dd5", "metadata": {}, "source": [ "User:
I've also spoken to my wife, and we decided that the pay hike isn't worth the misery. I'd rather live a little and take risks now, rather than chase a paycheck only to wake up one day too old to enjoy the fruits of my labor.
\n", "\n", "Model: \n", "
\n", "Sounds like you've got your answer, then — and it's telling that the two people whose judgment you actually trust (Joshua and your wife) both landed on the same side of it. Walking away from $500k to protect your agency and take a real swing at upside is exactly the \"skin in the game\" move you said you wanted.\n", "The one remaining box to tick is diligence on Nova AI itself: burn rate and runway at current growth, what's driving the losses (growth spend vs. structural margin problems), and the terms of your equity. Want to work through any of that?\n", "
" ] }, { "cell_type": "markdown", "id": "45730a17-0e0e-43bd-9be9-da53e34a1ada", "metadata": {}, "source": [ "(The model understood the question, brought up the $500,000 specifics, and even remembered niche expressions that the user mentioned previously like \"Skin in the game\", while continuing the previous conversation fluidly)" ] }, { "cell_type": "markdown", "id": "9acb1a17-96b4-41aa-aee4-53718ddf7354", "metadata": {}, "source": [ "## Everything that Broke, and learnings" ] }, { "cell_type": "markdown", "id": "f6ba2a01-10c2-4521-a2ed-8b9c30b8be80", "metadata": {}, "source": [ "### The choice of extractor model" ] }, { "cell_type": "markdown", "id": "8473e156-abd6-4cad-a677-04104105e178", "metadata": {}, "source": [ "During the first few rounds of graph construction, I used a lightweight model without reasoning capabilites in an attempt to make the extraction as token efficient as possible. This was mistake. With the model unable to reason and deliberate as to what information really deserved to make it into the graph, and double check its work, the quality of the graph produced was poor, with important user specific information never making into the graph, edges being invented for facts that should have been properties of an existing edge, and absence of second order edges between two nodes that were connected to the user node, but also connected to each other. \n", "\n", "On swapping the model for a reasoning model, the quality of output for smaller conversations went up immediately. The model identified all important facts, thought through higher order relationships, and also differentiated between information that was user-specific (and worth adding to the graph), vs universal facts that are known by LLMs already. However for longer conversations, the model thought for large amounts of time — sometimes up to 15 minutes and timed out without any output. One reason behind this was me sending the entire conversation transcript to the extraction model, and including the agent's turns, to provide context around why the user responded the way that they did. As the transcript got very long, the bookkeeping of facts that the model had to do became overwhelming, and it simply could not construct the graph in time. \n", "\n", "However, even after sending only the user's turns to the model, graph construction still took a long time (5-7) minutes before a suboptimal graph was produced. Looking at the reasoning logs of the model, I discovered that it essentially tried to build up the entire graph in its \"head\" (its reasoning trace) before commiting the entire graph all at once to the database. To do so, it would keep repeating the entire graph's schema every time it updated it, modifying it in its head as it made changes. The solution to this problem was to instruct the model to write every node and edge into the graph the moment it considered them, and modifying or appending to them in the graph later if new information about them was revealed later on in the conversation. \n", "\n", "This approach provided an extra layer of redundancy. If the model was cut off midway, its work wouldnt be lost, and it would be able to read the graph and pick up from where it left off when resumed. Both of these changes reduced extraction time for a reasonable conversation (with 10-20 turns) to around 2-3 minutes which was far more reasonable." ] }, { "cell_type": "markdown", "id": "77c9bb0e-3f53-4297-83c7-c98d16087bde", "metadata": {}, "source": [ "### Overfitting to the system prompt" ] }, { "cell_type": "markdown", "id": "4dd5117b-8afd-452f-a044-17e534f4f284", "metadata": {}, "source": [ "Another mistake that I made was including examples of previous extraction failures in the system prompt of the extractor model to demonstrate common failure modes and get the model to avoid them. This didn't work, and caused the model to overfit to the system prompt. It produced the right output when the exact example mentioned in the system prompt was tested, but failed when a different version of the same example was tested. \n", "\n", "However, by generalizing the instructions to the model, and having it reason over whether it was following those instructions, the quality of output got a lot better in general." ] }, { "cell_type": "markdown", "id": "78b103a4-607f-4dcd-b57d-dfe4b805d30d", "metadata": {}, "source": [ "### Attempting to build the model harness myself" ] }, { "cell_type": "markdown", "id": "d8d5429b-59c6-4fa5-a497-94d16d0fc354", "metadata": {}, "source": [ "To prototype early iterations of this system, I figured that if I couldn't reliably connect and test this memory system with chatbots on the web, it might make more sense to build a simple model harness locally on my laptop. I built one and used a free Gemini API key from Google AI Studio to set up a simple chat interface. This was not a great idea in hindsight. I wasted a lot of time building the harness and being unsatisfied with the output. To make things worse, I tightly coupled the memory system and harness together, to a point where I was spending more time trying to fix bugs in the harness than actually iterating on the memory system. At times, it was hard to locate the source of errors. Was there a mistake in how I built the harness, or how I connected it to the memory system, or in some part of the memory system itself? This was unclear.\n", "\n", "Eventually, I switched to Opencode, which should have been my first pick. Being open source, I was free to tinker with it as I liked. Fortunately, I didn't have to, and was able to build the memory system to be modular and work as a plugin with Opencode rather than requiring any deep integration. This also made it easier for others to try the system out as it reduced installation times, and was somewhat straightforward to setup as a plugin." ] }, { "cell_type": "markdown", "id": "5c0f8079-5788-4eea-a08a-c2df8e1be8eb", "metadata": {}, "source": [ "## Open Problems" ] }, { "cell_type": "markdown", "id": "bdbf9b1b-bf1e-464c-8a91-d6563b5cc093", "metadata": {}, "source": [ "### Extraction is still an active process" ] }, { "cell_type": "markdown", "id": "dd23c6db-a3bd-45f9-a972-42ac9068be78", "metadata": {}, "source": [ "The central principle behind building this system was to use the dynamics of memory in humans as a design inspiration for our architecuture. A key deviation from this principle however, is that the consolidation of memories into the graph in this system is still an active process that uses another LLM, unlike the passive consolidation that happens in humans — presumably when we sleep. \n", "\n", "I don't fully understand how consolidation of memories works in human brains, but I believe that it is pretty evident that we don't reason about our day when we sleep and chose to write things into our memory. Using an LLM to build the knowledge graph and decide what goes in and goes out is akin to that. \n", "\n", "Perhaps, in later implementations of this system, we will be able to connect the consolidation process more closely to the transformer running in the chat session itself, marking facts as important and worth remembering during the chat session rather than after it. Possibly, we may even be able to drop the knowledge graph with its engineered decay rates and artifical construction altogether, in favour of something even more natural, like a dedicated neural network that is small enough to learn important details continously on its own and act as a storehouse of information. In that context, phenomena like catastophic forgetting that is common with continual learning may be viewed as a feature rather than a bug, similar to how humans forget old memories to make room for new ones. However at this point these are all speculations, and I have no experimental backing to suggest that such an architecture is feasible." ] }, { "cell_type": "markdown", "id": "a589185a-5a60-4911-8458-afe24fc54eb5", "metadata": {}, "source": [ "### The system is currently untested at very large sizes" ] }, { "cell_type": "markdown", "id": "3d5a50ee-9f56-451c-b950-8b3f97debd4d", "metadata": {}, "source": [ "As of my writing of this writeup in September 2026, I have not yet been able to test the graph at extremely large sizes and observe its behaviour when holding extremely large amounts of information. It shows promise compared to sumary based memory systems, and recalls specific details far more accurately than providers on the internet when tested over a small number of conversations (10 - 15 turns over 4-5 conversations), but unexpected behaviour is possible at larger sizes. I plan on writing a follow up to this writeup in the near future after I have stress tested the system for its failure modes." ] }, { "cell_type": "markdown", "id": "951e3eff-2599-495c-9b4e-06bfa3ead70b", "metadata": {}, "source": [ "### Gimmicks with the Opencode Plugin" ] }, { "cell_type": "markdown", "id": "9718a350-6379-4440-8a58-a48f9a4a911c", "metadata": {}, "source": [ "The opencode plugin for this system is registered on the npm registry under the name \"pegasus-opencode\". The current implementation of this plugin (version 0.2.0) has a few issues that need to be worked out, such as the first write to the graph often taking very long (3-4 minutes), and possibly some other issues that I haven't hit yet. Working on resolving these as quickly as possible." ] }, { "cell_type": "markdown", "id": "80ee370f-f21b-4a46-a833-f8abdb771cb9", "metadata": {}, "source": [ "## Trying it yourself" ] }, { "cell_type": "markdown", "id": "03b82d73-4611-44d9-9429-589d33d573ae", "metadata": {}, "source": [ "Since the nodes and edges of the graph resemble a constellation of stars, I've decided to call this system \"Pegasus\", although in most parts of the code, you will simple see it referred to as \"onto-mem\" or some variation of that name." ] }, { "cell_type": "markdown", "id": "53340112-e86a-4de6-bc12-55a69b7eefef", "metadata": {}, "source": [ "The entire source code for this project is open-source with an MIT license at my Github: [Pegasus - Github](https://github.com/Stponnada/Pegasus)" ] }, { "cell_type": "markdown", "id": "15ceba42-fe41-449b-bbd3-93e199d6ce20", "metadata": {}, "source": [ "Currently, this engine as a plugin only exists for Opencode, (preferrably with a go subscription) but I am working on making it available for Claude Code and Codex as well. \n", "\n", "To install for Opencode, run:\n", "\n", "```bash\n", "npm install pegasus-opencode\n", "opencode plugin pegasus-opencode\n", "```" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.9.6" } }, "nbformat": 4, "nbformat_minor": 5 }