A portable method for turning an LLM's memory of you into a structured graph — one that separates observations from interpretations, flags which claims are load-bearing, and surfaces insights that only appear at the intersection of multiple patterns.
Built on Patrick McCarthy's open-knowledge-graph schema, adapted for personal memory rather than scientific claims.
Narrative companion essay: Know Thyself: a schema for personal memory in LLM conversations.
After a while of using Claude, your memory accumulates. You end up with a flat list of claims — some rock-solid ("born in Berlin"), some interpretive ("I stay in misaligned situations because of X"), some repeated so many times they start to feel settled even though they rest on a single inference. The flat list flattens these. A claim stated once feels the same as a claim confirmed by five independent events.
This scaffold restructures that flat list into a graph with explicit node types (fact, episode, pattern, interpretation, intersection-produced insight, external-theory bridge, open question, and — optionally — operating rules derived from those patterns) where every claim carries its provenance and can be checked.
The invariant, from Pat McCarthy's work: a claim without provenance is indistinguishable from noise. Every node and every edge must carry (attribution, evidence, derivation) — who stated it, what it rests on, how it was derived.
The operational rule: attribution ≠ confidence. A claim you've restated across five conversations isn't five pieces of evidence; it's one derivation repeated five times. Real confidence accumulates only from independent derivations — different episodes, different contexts, different evidence types.
At the end of the process:
- A YAML graph file with typed nodes and edges, fully traceable provenance
- A visual diagram (PDF + PNG) you can print and keep somewhere physical
- A "load-bearing" list — the observations most of your interpretations rest on
- A "fragile" list — the interpretations flagged tentative, with explicit caveats
- An ongoing method for integrating new events into the graph as they happen
The graph is operational, not therapeutic. It helps you see your own patterns more clearly. It is not a diagnosis, not a treatment plan, not a substitute for talking to a human who knows you.
- Read
SAFETY.mdfirst. Five minutes. Important. - Open a Claude conversation where you already have meaningful memory accumulated. (If you don't, this scaffold will be thin; come back after a month or two of real use.)
- Paste the contents of
START_HERE.mdinto that conversation. - Claude will walk you through the construction. Expect it to take 20–45 minutes of back-and-forth. Push back on anything that doesn't fit.
- When Claude produces your YAML graph, save it. Use
render.pyto generate a visual, orprintable.pyfor a printable PDF. - Come back to the graph when new events happen. The graph should grow — cautiously, with new evidence explicitly tied to existing nodes.
| File | Purpose |
|---|---|
README.md |
This file |
SAFETY.md |
Important caveats — read first |
START_HERE.md |
The prompt to paste into Claude |
SCHEMA.md |
Formal specification of node types and edges |
RELATED_FRAMEWORKS.md |
Survey of adjacent schemas (PROV-O, Toulmin, Zettelkasten, epistemic status, PKG) and what this scaffold borrows from each |
example-graph.yaml |
A small fictional example showing the schema |
render.py |
Generate a graphviz diagram from your YAML |
printable.py |
Generate a multi-page printable PDF |
The underlying epistemic framework (confidence chains, provenance triples, emergent nodes from intersection) comes from Patrick McCarthy's open-knowledge-graph, released MIT. The adaptations for personal memory (observation as a first-class node type, type-tier confidence instead of a numeric score, HANDLING directives for sensitive content, natural-experiment evidence type, first-class open questions) are modifications you're free to modify further for your own use.
