Layered recall

Layered Recall

Progressive memory system with two orthogonal dimensions of lazy loading:

1. Scope layers - What is relevant (identity, project, domain, deep)
2. Depth layers - How much detail to fetch (IDs, summary, full)

Combined savings: 10-50x tokens vs eager loading.

Depth Pattern (Fetch-on-Confirmation)

Instead of loading full memory entries upfront, agents fetch in 3 depths:

Depth 1: IDs only        (~10 tokens per match)
  Agent decides which are worth investigating

Depth 2: Summary         (~50 tokens per match)
  Room, type, preview (first 80 chars)
  Agent confirms relevance

Depth 3: Full content    (~500+ tokens per match)
  Only fetched for confirmed matches

Example flow:

1. Agent searches "auth refresh token"
2. Depth 1 returns 8 IDs: d-abc123, d-def456, ...
3. Agent requests Depth 2 for IDs 1-3
4. Sees room=authentication, type=decision, preview="Chose JWT..."
5. Agent confirms IDs 1,3 are relevant
6. Requests Depth 3 only for those 2 entries
7. Gets full content for ~1000 tokens instead of 4000+

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.json filtered by domain

Trigger: Intent classifier detects domain (e.g., "fix the login bug" -> room: authentication)

Layer 4: Deep Search (~2-5K tokens, explicit)

Only loaded when explicitly needed or when Layers 1-3 don't have enough context. Contains:

• Full semantic search results
• Cross-project pattern matches
• Historical error resolutions
• Archived decisions

Source: PostgreSQL vector search + palace cross-wing search

Trigger: Agent explicitly queries, or user asks "have we done this before?"

Recall Flow

Session Start
  -> Load Layer 1 (identity)
  -> Detect project -> Load Layer 2 (facts)
  -> User sends prompt
  -> Classify intent/domain -> Load Layer 3 (room)
  -> If insufficient context -> Load Layer 4 (deep)

Token Budget

Layer	Tokens	When
L1	~200	Always
L2	~500	Per project
L3	~1-2K	Per task domain
L4	~2-5K	On demand
Total max	~8K	Worst case

vs. loading everything: ~30-50K tokens

Savings: 4-6x token reduction

Integration

With Existing Hooks

• instinct-loader -> feeds Layer 2 and Layer 3
• smart-memory-recall -> implements Layer 3 scoring
• intent-classifier -> triggers Layer 3 room selection
• graph-indexer -> powers Layer 4 deep search

With Memory Palace

• Layer 2 pulls from palace wing index
• Layer 3 pulls from palace room drawers
• Layer 4 searches across all wings

With Agents

• Agents inherit parent's Layer 1-2 context
• Each agent can request Layer 3-4 for their domain
• Agent memories feed back into palace for future recall