Anamnesis

Aristotelian Memory Architecture for AI Coding Agents

Anamnesis is a persistent memory system for AI coding agents, grounded in Aristotle's Posterior Analytics and his 5-layer epistemological hierarchy. It enables agents to learn from experience, retain causal knowledge, and improve over time.

Aristotelian Foundation

The system implements Aristotle's hierarchy of knowledge acquisition:

Layer	Greek	Role	In Anamnesis
Sophia	Wisdom	Unchanging first principles	Hardcoded safety axioms (rm -rf, force push)
Episteme	Scientific Knowledge	Justified, causal, universal	Promoted nodes: confidence >= 0.85, helpful >= 3, supports >= 2
Techne	Craft Knowledge	Practical know-how	Reflection-generated hypotheses with WHEN/THEN/BECAUSE structure
Empeiria	Experience	Accumulated observations	Tool events, errors, successes recorded per session
Aisthesis	Perception	Raw sense data	Real-time tool invocations and outputs

Architecture

Session Start ──> Briefing (relevant knowledge surfaced)
      │
Prompt Submit ──> Retrieval (domain-filtered, threshold-gated)
      │
Tool Execution ──> Observer (records empeiria)
      │
Tool Result ──> Feedback (confirm/refute surfaced knowledge)
      │
Session End ──> Reflection Pipeline (5-step LLM: Narrate→Diagnose→Hypothesize→Validate→Store)
      │
      └──> Knowledge Graph (nodes + edges + decay + promotion)

Key Components

• v2/db.py -- SQLite schema with FTS5 full-text search, domain classification, node lifecycle
• v2/retrieval.py -- Dual-process retrieval: FTS5 (System 1) + Graph expansion (System 2) + adaptive threshold
• v2/reflection.py -- 5-step LLM pipeline with quality gating (salience scoring)
• v2/knowledge_graph.py -- Graph operations: BFS traversal, promotion, Ebbinghaus decay
• v2/feedback.py -- Cosine-similarity credit assignment, confidence updates
• Hook entry points -- on_session_start.py, on_prompt_submit.py, on_post_tool.py, on_pre_bash.py

v3 Features (Retrieval Precision)

Addressing the Doctrine of the Mean -- too little knowledge = baseline, too much = noise:

1. Domain Genus Classification -- 8 domains (web_backend, data_processing, database, etc.) prevent cross-domain retrieval pollution
2. Adaptive Relevance Threshold -- Ratio-based filtering that tightens as the knowledge graph grows
3. Write-Time Quality Gating -- Salience scoring (novelty + specificity + actionability) rejects platitudes
4. Enhanced Deduplication -- Jaccard 0.8 threshold, error-family-aware at 0.65

Experiment Results

Phase	Description	Turns	Reduction
P1	Baseline (no memory)	554	--
P2	Anamnesis v2 (basic)	426	-23.1%
P3	Deep Knowledge (WHEN/THEN/BECAUSE)	489	-11.7%
P4	v3 (domain + threshold + quality gate)	TBD	TBD

25-task benchmark across 5 categories (Flask microservices, CSV pipelines, config apps, SQLite migrations, API clients).

Test Suite

146 tests across 11 test files:

cd /path/to/anamnesis && python3 -m pytest v2/tests/ -v

Installation (Claude Code Hooks)

Configure in ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [{"matcher": "Bash", "hooks": [{"type": "command", "command": "python3 /path/to/anamnesis/on_pre_bash.py"}]}],
    "PostToolUse": [{"matcher": "", "hooks": [{"type": "command", "command": "python3 /path/to/anamnesis/on_post_tool.py"}]}],
    "PrePromptSubmit": [{"matcher": "", "hooks": [{"type": "command", "command": "python3 /path/to/anamnesis/on_prompt_submit.py"}]}],
    "SessionStart": [{"matcher": "", "hooks": [{"type": "command", "command": "python3 /path/to/anamnesis/on_session_start.py"}]}],
    "Stop": [{"matcher": "", "hooks": [{"type": "command", "command": "python3 /path/to/anamnesis/agent_learn.py"}]}]
  }
}

Research

• Novel application of Aristotle's Posterior Analytics to ML memory architecture (zero prior art found in literature review)
• Complementary Learning Systems theory (hippocampal fast binding + cortical slow consolidation)
• Informed by: ExpeL, CLIN, G-Memory, SEEKR, LightRAG, Karpathy's LLM Wiki

See docs/anamnesis_v3_brainstorm.md for the full literature review and design rationale.

License

MIT