Memory Palace

Hierarchical, persistent memory system for Claude Code sessions. Organizes knowledge so agents never lose context across sessions.

Architecture

Palace (global)
  Wing: project-name
    Room: authentication
      Drawer: "chose JWT over sessions" (2026-04-07)
      Drawer: "rate limiting at 100 req/min" (2026-04-06)
    Room: database
      Drawer: "PostgreSQL with pgvector" (2026-04-05)
      Drawer: "migration strategy: blue-green" (2026-04-04)
    Room: deployment
      Drawer: "Vercel + GitHub Actions" (2026-04-03)
  Wing: another-project
    Room: ...

Storage Format

All memories stored in ~/.claude/palace/ as flat JSONL files per wing:

~/.claude/palace/
  index.json          # Wing registry
  my-project.jsonl    # All drawers for this wing
  other-project.jsonl # All drawers for this wing

Drawer Entry Format

{
  "id": "d-abc123",
  "wing": "my-project",
  "room": "authentication",
  "content": "Chose JWT with refresh tokens over session-based auth. Reason: stateless, mobile-friendly, scalable.",
  "tags": ["auth", "jwt", "architecture"],
  "timestamp": "2026-04-07T14:30:00Z",
  "session_id": "s-xyz789",
  "agent": "architect",
  "type": "decision"
}

Entry Types

• decision - Architectural or design choice with reasoning
• discovery - Something learned about the codebase
• error - Error encountered and how it was resolved
• constraint - External limitation or requirement
• pattern - Recurring code or workflow pattern

Operations

Store a Memory

When an important decision, discovery, or constraint is identified:

Wing: detect from CLAUDE_PROJECT_DIR or package.json name
Room: infer from topic (auth, db, deploy, frontend, api, etc.)
Content: what was decided/discovered and WHY
Tags: relevant keywords for search

Recall Memories

Before starting work on a topic, query relevant rooms:

1. Get wing from current project
2. List rooms with recent activity
3. Load drawers matching current task keywords
4. Inject as context: "Previous decisions in this area: ..."

Search Across Wings

For cross-project pattern detection:

Search all wings for entries matching query
Return sorted by relevance (keyword match + recency)
Useful for: "Have I solved this before in another project?"

Integration Points

With Agents

• architect: Store architectural decisions automatically
• self-learner: Store error resolutions as discoveries
• planner: Load relevant rooms before planning
• compass: Use palace for session recovery context

With Hooks

• palace-auto-save hook: Captures decisions from conversation
• palace-recall hook: Injects relevant memories at session start
• pre-compact-continuity: Saves WIP state to palace before compaction

With Existing Memory System

Palace complements (does not replace) the existing PostgreSQL memory:

• PostgreSQL: raw embeddings, vector search, learning store
• Palace: structured, human-readable, hierarchical organization
• Both can be queried; palace is faster for targeted recall

Room Detection Heuristics

When to Store

ALWAYS store:

• Architectural decisions with reasoning
• External constraints (client requirements, hosting limits)
• Error resolutions that took > 5 minutes to solve
• Chosen patterns and why alternatives were rejected

NEVER store:

• Code snippets (they're in git)
• Temporary debugging notes
• Information already in CLAUDE.md
• Ephemeral task progress (use tasks for that)