Turn an LLM's memory of you into a structured graph that knows what it knows — and what it's just guessing.

A flat memory list treats a claim repeated five times as five pieces of evidence. It isn't. This scaffold restructures that list into typed nodes — fact, episode, pattern, interpretation, open question — where every claim carries provenance (attribution, evidence, derivation) and confidence accumulates only from independent derivations.

Companion essay: Know Thyself — the full argument lives there.

Claim: Restructuring flat LLM memory into a provenance-typed graph yields honest confidence accumulation. Grounds: Stated thesis of the scaffold; elaborated in the companion essay. Status: stipulated Leans on: the schema and prompt below; SCHEMA.md for the typed-node spec.

Quickstart — how to run the scaffold end-to-end

1. Read SAFETY.md (5 minutes).
2. In a Claude conversation that has accumulated real memory, paste START_HERE.md. (Claude Code: a /know-thyself slash command does the same thing.)
3. Save the YAML Claude produces. Render it:

pip install pyyaml graphviz
python render_dashboard.py your-graph.yaml   # interactive HTML, NOW node centered

render.py builds a static graphviz diagram, render_mandala.py does concentric rings, printable.py builds a multi-page PDF. Each script tells you which pip install it needs.

Claim: Three steps (safety read → paste prompt → render) are sufficient to produce a usable graph. Grounds: Operational instructions; verified by the bundled renderers. Status: stipulated Leans on: START_HERE.md (the prompt), SCHEMA.md (the YAML shape), the render scripts in this repo.

What's here — the file inventory of this repo

Claim: These files are the complete public surface of the scaffold. Grounds: Direct enumeration of the repo contents. Status: stipulated Leans on: every cross-reference elsewhere in the README and START_HERE.md.

What I built — the structural extensions for personal memory

I built a personal-graph schema: typed nodes for a single life, four-scale confidence synthesis, temporal validity, MCP retrieval. The provenance-triple shape underneath is older — RDF and PROV-O ship it as W3C standards, and Patrick McCarthy's open-knowledge-graph gives a contemporary articulation with formal necessity arguments for the scientific case. The structural extensions for personal memory below are mine:

1. Observation as a first-class node type. In a scientific graph, observations recede after grounding a proposition. In a personal graph they get reinterpreted — the first three months mean one thing in November and another in May. Keeping the episode separate stops interpretations collapsing back into the events that generated them.
2. A valid_at axis. Propositions about a person aren't permanently valid the way physical-law propositions are; every claim carries a validity window that decays unless re-grounded.
3. Inverted edge-density prediction. Mature scientific graphs become edge-dense. Personal graphs don't — new life events spawn new nodes, cross-time edges stay sparse. A mature personal graph is node-dense with sparse adjacency.

Smaller extensions: type-tier confidence (no replication, no external ground truth, so a numeric score is dishonest); HANDLING directives for sensitive content; natural-experiment evidence type for life events that function like A/B tests; open questions as first-class nodes.

Claim: The personal-graph schema is a deliberate set of structural extensions on top of an older provenance-triple shape. Grounds: Builder's account; cross-checked against PROV-O / RDF / open-knowledge-graph as the prior art. Status: stipulated Leans on: SCHEMA.md (formal spec of the extensions), the Credit section below.

Credit — where the underlying ideas come from

• W3C RDF (2004) and PROV-O (2013) — the typed-triplet shape as web standards.
• George Miller (1956) and Nelson Cowan (2001) — working-memory bounds that motivate why a graph beats a flat list.
• Patrick McCarthy's open-knowledge-graph — formal necessity arguments for the scientific case. McCarthy's repo is unlicensed at time of writing; the schema as a structural taxonomy is treated here as uncopyrightable, and this repository's MIT license covers the specific implementation, prose, and renderers — not the underlying ideas.

Adjacent prior art (Anthropic's citations API, Park et al.'s Generative Agents, MCP Knowledge Graph Memory, Mem0, Graphiti, and others) is surveyed in the companion essay.

Claim: The provenance-triple lineage is W3C-standardized; the personal-graph extensions are the new contribution here. Grounds: Cited W3C standards (RDF 2004, PROV-O 2013), Miller (1956), Cowan (2001), McCarthy's open-knowledge-graph repo. Status: established (for the cited prior art); stipulated (for the boundary between borrowed shape and new extensions) Leans on: RELATED_FRAMEWORKS.md for the broader survey; the companion essay for adjacent prior art.