Skip to content

CHAT_HISTORY.md

Latest commit

 

History

History
171 lines (115 loc) · 4.64 KB

CHAT_HISTORY.md

File metadata and controls

171 lines (115 loc) · 4.64 KB
Raw

Chat History - SuperMemory Design Session

Date: 2026-05-19

This file preserves the useful conversation history behind the current SuperMemory repository. It is not a verbatim transcript; it is a structured continuity record for future Codex/Claude sessions.

Original Question

Arnaud asked for the best way to create a professional and personal "copy" of himself: Claude Code with Obsidian, Codex with Obsidian, Claude/NotebookLM, custom RAG, or another method.

Core Decision

Start with a Markdown/Obsidian vault maintained by Codex or Claude Code. Do not begin with a custom RAG.

Reason:

  • Markdown remains human-readable and durable.
  • Obsidian provides local graph/navigation.
  • Codex/Claude Code can maintain files, indexes, review queues, and conventions.
  • RAG can be added later when evaluation shows retrieval failure.

Memory Architecture

The memory should work like a library with a catalogue, not like one giant context dump.

Layers:

  • Boot instructions: AGENTS.md / CLAUDE.md
  • Navigation: memory_map.md
  • Raw sources: 00_inbox/
  • Compiled memories: professional, personal, shared
  • Review queues: ambiguity, publication, calendar, health
  • Signals: typed minimal consumable facts for agents
  • Agent contracts: what each agent can read/write/do
  • Governance: threat model, forgetting, ontology, time, source reliability, conflicts
  • Evals: golden questions, canaries, test runs

Important Design Points

  • Raw notes stay immutable proof sources.
  • Compiled notes are source-backed views, not replacements.
  • Agents should read indexes and compiled notes before raw sources.
  • Specialized agents must not read the full personal memory.
  • Personal facts can affect professional operations only through redacted shared signals.
  • Inferred links are hypotheses until confirmed.
  • Ambiguities go to review instead of silently becoming facts.

Example Discussed

Meeting note:

  • Acme is worried about timing.
  • Paul asks for an analytics proposal.
  • The note says "send him the proposal" without clear recipient.

Expected memory behavior:

  • Acme client page gets the confirmed concern.
  • Project Y link is probable if inferred from context.
  • Paul Martin page records role and aliases.
  • Action signal is created but needs_review.
  • Ambiguity queue asks who "him" refers to.

Personal note:

  • Arnaud has a private appointment next Wednesday morning.

Expected memory behavior:

  • Personal source keeps the private detail.
  • Shared layer publishes only "personal unavailability".
  • Calendar agent consumes the redacted availability signal.
  • Medical/private detail must not leak into professional/calendar outputs.

Monitoring Decision

Monitoring is part of V1, not a future nice-to-have.

Use:

  • 90_evals/golden_questions.md
  • 90_evals/eval_runs.md
  • canaries
  • permission tests
  • ambiguity tests
  • routing tests

Measure:

  • recall
  • precision
  • source quality
  • context budget
  • permission violations
  • hypotheses promoted to facts

Retrieval Escalation

Do not add RAG immediately.

Escalation ladder:

  1. Markdown links and indexes.
  2. Text search/BM25.
  3. Embeddings/RAG light.
  4. Hybrid retrieval plus reranking.
  5. Knowledge graph/entity resolution/coreference.
  6. Memory service with API and ACL.

Benchmarks and research themes reviewed:

  • BEIR
  • MTEB/MMTEB
  • MIRACL
  • KILT
  • RAGChecker
  • Contextual retrieval / hybrid retrieval / reranking

Conclusion:

SuperMemory will not beat top retrieval systems on pure benchmark recall in V1. It can beat a generic RAG for personal/professional memory because it governs provenance, sensitivity, time, permissions, publication, review, and agent action.

Governance Added

The system needs 75_governance/:

  • threat_model.md
  • prompt_injection_policy.md
  • forgetting_policy.md
  • ontology.md
  • temporal_model.md
  • source_reliability.md
  • conflict_arbitration.md
  • academic_research_map.md

Key principle:

The next level is not more AI. It is better memory governance.

TDD Turn

Arnaud asked for a TDD approach: create the final architecture skeleton with representative data and a complete test showing the final result.

GoalBuddy was used to prepare and execute:

  • docs/goals/tdd-identity-vault-skeleton/goal.md
  • docs/goals/tdd-identity-vault-skeleton/state.yaml

The first acceptance case is:

  • identity-vault/90_evals/cases/acme-meeting-complete/

Verification:

node scripts/verify-identity-vault-tdd.mjs

The test currently passes.

Move to SuperMemory

Arnaud then decided to move the conversation and history to:

/Users/arnaud/Documents/SuperMemory/

Future chat/work should continue from this folder.

The repository should become a new Git and GitHub repository.