layered-recall

# Layered Recall 4-layer progressive memory system. Each layer adds more context only when needed, saving tokens while ensuring nothing important is missed. ## The 4 Layers ``` Layer 1: Identity (always loaded, ~200 tokens) Who is the user? What are their preferences? Layer 2: Critical Facts (per-project, ~500 tokens) Hard constraints, active decisions, blockers Layer 3: Room Recall (on-demand, ~1-2K tokens) Relevant memories for current task domain Layer 4: Deep Search (when needed, ~2-5K tokens) Full semantic search across all memories ``` ## Layer Details ### Layer 1: Identity (~200 tokens, ALWAYS loaded) Loaded at every session start. Contains: - User preferences (language, style, autonomy level) - Global constraints (no emojis, Turkish responses, etc.) - Tool preferences (which editors, which terminal) **Source:** `~/.claude/projects/*/memory/user_*.md` ### Layer 2: Critical Facts (~500 tokens, per-project) Loaded when entering a project directory. Contains: - Active architectural decisions - Known blockers and constraints - Current sprint/milestone goals - Tech stack and versions **Source:** `~/.claude/projects/*/memory/project_*.md` + `thoughts/CONTEXT.md` ### Layer 3: Room Recall (~1-2K tokens, on-demand) Loaded when task domain is detected (auth, database, deploy, etc.). Contains: - Previous decisions in this domain - Past errors and fixes - Patterns that worked - Patterns that failed **Source:** Memory palace rooms + `mature-instincts.jso...