Memory System

Overview

TARS memory operates across three layers: file-based narrative storage, Postgres structured data, and the wiki as the curated knowledge layer.

Architecture

Session History (raw)
  ~/.claude/projects/**/*.jsonl (1,266+ files, 177MB)
      │
      ▼ [learning-synthesis.ts hook at SessionEnd]
      │
  ┌───┴───────────────────────────┐
  │                               │
  ▼                               ▼
MEMORY/Learning/{PHASE}/     Postgres learnings table
  (markdown narrative)        (structured + embeddings)
  65+ files                   57+ rows
      │                               │
      ▼                               ▼
  Wiki (curated)              Postgres signals table
  wiki.tedlango.com           (ratings, failures, patterns)
  (published knowledge)       117+ rows

File-Based Memory (~/.claude/MEMORY/)

Learning Phases

• OBSERVE — Context gathering insights
• THINK — Option generation patterns
• PLAN — Approach selection learnings
• BUILD — Implementation patterns
• EXECUTE — Execution discoveries
• VERIFY — Testing and validation insights
• LEARN — Meta-learnings and corrections

Learning Pipeline

The learning pipeline runs at SessionEnd and processes the session’s JSONL transcript:

1. Classify — Deterministic: count turns, tool calls, detect project. Skip trivial sessions.
2. Extract — Heuristic: find corrections, error→fix patterns, decisions, discoveries.
3. Write — Dual-write: markdown to MEMORY/Learning/ + row to Postgres with tags.
4. Signal — Record quality signals (failures, patterns) to Postgres.

TELOS (Life Context)

Deep personal context at ~/pai/TELOS/:

• MISSION.md, GOALS.md, PROJECTS.md, BELIEFS.md
• MODELS.md, STRATEGIES.md, NARRATIVES.md
• LEARNED.md, CHALLENGES.md, IDEAS.md

Loaded by skills that need personal context (PAIUpgrade, JobSearch, etc.).