Knowledge dashboard
A self was never flat
The essay's knowledge, dashboard-shaped — graph, eyes, actions. The graph is the source of truth; the prose follows.
The essay's claims, typed and connected. Synthesis is the author's contribution. Citation grounds in a named source. Derivation follows from prior nodes. The spine — graph IS the memory — sits at center; everything orbits it. Click any node to jump to its card.
S S01-mirror-problem · opening (Alex case)
Flat memory turns repetition into evidence — the model speaks back to the user as a mirror, sycophancy hardened into self-knowledge.
C C06-mirror-study
Condensed user-profile memory produced the largest sycophancy amplification of any personalization feature studied — measured effect, not metaphor.
grounds: MIT & Penn State, Personalization features can make LLMs more agreeable (CHI 2026)
S S02-fix-is-shape · turn (after Alex)
The fix is not more memory. The fix is shape.
derives_from: S01-mirror-problem
C C09-karpathy-wiki
LLM Wiki — plain markdown, model self-edits, lint loop catches contradictions and orphans. No types, no provenance.
S S03-memory-must-have-types · the shape
Memory must have types — episodic and semantic held in distinct stores, not collapsed into a flat list.
derives_from: S02-fix-is-shape
C C04-tulving
Episodic vs semantic memory as separate stores — the binding principle the schema operationalizes.
grounds: Tulving, Episodic and Semantic Memory (1972)
C C05-coala
Working / episodic / semantic / procedural — the canonical agent-memory taxonomy. The eight node types here are an opinionated refinement of the semantic side.
grounds: Sumers, Yao, Narasimhan, Cognitive Architectures for Language Agents (2023)
C C12-mccarthy-edges
Edge vocabulary — derives_from, evidences, grounds, overlaps_with, generalizes — each carrying an (attribution, evidence, derivation) triple.
S S07-attribution-neq-confidence · the rule
Attribution ≠ confidence. The operating rule.
derives_from: S03-memory-must-have-types , C12-mccarthy-edges
S S08-repetition-isnt-corroboration · the rule, applied
Six conversations saying the same thing is one derivation repeated six times, not six pieces of evidence. A novel cannot quietly become an overlap; it waits for a new, independent observation.
derives_from: S07-attribution-neq-confidence
C C10-memanto
Typed semantic memory with information-theoretic retrieval — thirteen vector categories, no graph; 89.8% on LongMemEval, 87.1% on LoCoMo, beating graph-hybrid baselines on QA recall.
C C11-anthropic-memory
Client-side persistence exposed as a /memories directory the model can view / create / str_replace / insert / delete / rename.
grounds: Anthropic memory tool (Apr 8 2026)
S S13-graph-is-the-memory · the punchline
The graph IS the memory. A YAML file the user owns; models read it and pick up the thread; when a model retires, the graph stays.
derives_from: C12-mccarthy-edges , C11-anthropic-memory
Claims that need more grounding. The essay's confident synthesis — but where external corroboration would strengthen the case, or an open question is implicit.
The eight node types (Reference / Observation / Overlap / Novel / Emergent / Equivalency / Practice / Open) are the author's opinionated refinement of CoALA's four. The specific cuts — especially Novel, Emergent, Open — are author-original.
Why eyes: CoALA grounds working/episodic/semantic/procedural. The further partition of the semantic side is the essay's contribution, not yet validated externally. Empirical work showing these specific cuts reduce one-derivation-as-six errors would harden the case.
ground_with:
- CoALA — working/episodic/semantic/procedural
- Park et al., Generative Agents (UIST 2023) — observation vs reflection separation
- A targeted user study on typed-vs-flat memory and sycophancy outcomes
S08-repetition-isnt-corroboration
The Mira/Alex case is illustrative, not empirical. No user study cited testing whether graph-shaped memory reduces sycophancy at the user-experience level — only the CHI 2026 paper showing flat memory amplifies it.
Why eyes: The negative result (flat memory amplifies sycophancy) is grounded. The positive result (typed-and-edged memory reduces it) is asserted by construction, not measured. The essay would land harder with a paired study.
ground_with:
- CHI 2026 mirror study (the negative case, already cited)
- A within-subjects study comparing flat vs typed memory on sycophancy metrics
The architectural claim ("graph IS the memory") rests on two recent shipped/proposed pieces (Anthropic memory tool, McCarthy edge vocabulary). It is a synthesis claim, not yet stress-tested at scale.
Why eyes: The composition is sharp but new. Any production deployment story — even a single user running the YAML graph against a real model over months — would move this from elegant proposal to demonstrated architecture.
ground_with:
- Anthropic memory tool docs
- Stateful vs stateless agent architecture surveys
- A longitudinal case study of one user running know-thyself for ≥6 months
Memanto's benchmark wins (89.8% LongMemEval, 87.1% LoCoMo) are dispatched in one sentence: "the benchmark measures recall." That move needs more weight.
Why eyes: A reader sympathetic to vector-only memory will not be convinced by a one-line dismissal. An eyes-item that names the specific benchmark task (multi-hop fact retrieval) and contrasts it with the corroboration-provenance task the essay cares about would close the gap.
ground_with:
What the essay implies for the reader. Etudes drill specific claims; apply items put the principle to work; next points at sibling essays.
etude Six restatements →
One claim, said six ways. Watch the flat list count it as six. Watch the typed graph count it as one. Drills S08.
etude Attribution vs confidence →
Sort five claims about Alex by attribution count and by independent-derivation count. The two columns disagree. Drills S07.
etude Flat vs typed →
Same set of conversations, two memories. The flat memory reads back a confident self-diagnosis; the typed graph reads back a tentative novel. Demonstrates S02 and S13.
apply Audit your last LLM conversation
If memory was on, what claims about you did the model state confidently? For each: how many independent episodes ground it, and how many times did you simply restate it? The two counts should not be conflated.
apply Find your spine
Of the things you believe about yourself, which three or four observations are load-bearing — referenced by the most downstream claims? Miscoding one of those cascades. Knowing which is which is the work.
apply Name your novels
Pick three things you have been quietly believing about yourself. For each, write the single episode it derives from. If you cannot find a second independent episode, the claim is a Novel, not an Overlap. Mark it tentative.
An audit of this essay against the same schema it argues for. Five dimensions: grounded (load-bearing claims cite real evidence), tentatives (speculative claims marked tentative), synthesis (credit lands honestly), provenance (the reader can verify), voice (claim leads, no hedging). Composite is the read across them.
Tight on every dimension. MIT/Penn State CHI 2026 sycophancy study, CoALA, Karpathy's LLM Wiki, Memanto, McCarthy's edge vocabulary, Tulving, Anthropic memory tool, Park et al — citations are real and load-bearing. The Novel type is itself the embodiment of marking-tentative; the schema enacts its own discipline.
Soft spot · repair
the opening claim — "a model's memory is a list — bullets, facts" — generalizes from ChatGPT-style memory to all LLM memory. False of Memanto, Mem0, Graphiti, A-Mem. The argument doesn't need the generalization. Specify "ChatGPT's persistent memory" and the precision rises without losing the move.