Unified context-gathering and learning system for Claude Code. Merges the best parts of cc-dejavu, homunculus, claude-mem, and serena into a single SQLite-based tool with frecency ranking and instinct learning.

• Session Sync - Indexes your Claude Code session history (JSONL files)
• Command Extraction - Every bash command with description, output, and working directory
• Pattern Detection - Automatically creates "instincts" (learned behaviors) from Claude's thinking
• Frecency Search - Full-text search with BM25 ranking + frequency x recency scoring
• Context Injection - Hooks for Claude Code to inject relevant context into prompts
• Telemetry - Track context hit rates and latency percentiles
• Evidence-Based Validation - External signals (tests, builds, commits) validate patterns
• Durability - WAL mode, snapshots, and integrity checks

# Navigate to the project
cd ~/.claude/claude-memory-system
# Install dependencies (via WSL on Windows)
wsl -e bash -lc "cd ~/.claude/claude-memory-system && bun install"
# Sync your Claude Code sessions
wsl -e bash -lc "cd ~/.claude/claude-memory-system && bun run src/cli.ts sync"
# View status
wsl -e bash -lc "cd ~/.claude/claude-memory-system && bun run src/cli.ts status"
# Search for commands
wsl -e bash -lc "cd ~/.claude/claude-memory-system && bun run src/cli.ts search 'docker'"

-h, --help Show help message
-f, --force Force full resync (ignore byte offsets)
-t, --type Filter by type: command, observation, instinct, all
-p, --project Filter by project path
-n, --limit Maximum results (default: 20)
-d, --domain Filter instincts by domain
-j, --json Output as JSON
--days Days for decay threshold (default: 30)
--hours Hours for telemetry window (default: 24)

Every bash command Claude ran, with:

• The command itself
• Description (what Claude said it does)
• Working directory
• Output (stdout/stderr)
• Timestamp
• Frequency count

Patterns detected from Claude's thinking blocks:

Context retrieval performance metrics:

• Hit rate (context found vs missed)
• Latency percentiles (P50, P90, P99)
• Timeout tracking (>100ms)

~/.claude/claude-memory-system/
├── src/
│ ├── cli.ts # CLI entry point
│ ├── db/
│ │ ├── schema.ts # Database tables and types
│ │ ├── connection.ts # SQLite connection + durability
│ │ ├── frecency.ts # Frecency ranking algorithm
│ │ └── telemetry.ts # Telemetry recording
│ ├── sync/
│ │ └── sync.ts # Session file syncing
│ ├── search/
│ │ └── search.ts # FTS5 search with BM25
│ ├── patterns/
│ │ └── detector.ts # Pattern detection from thinking
│ ├── evolution/
│ │ └── reinforcement.ts # Decay and pruning
│ └── hooks/
│ └── context.ts # Claude Code hook handlers
├── package.json
└── tsconfig.json
~/.claude-memory-system/
├── memory.db # SQLite database (auto-created)
└── memory-snapshot-*.db # Backup snapshots

• sessions - Claude Code session metadata
• commands - Extracted bash commands with FTS5
• observations - Thinking block extracts with FTS5
• instincts - Learned patterns with FTS5
• indexed_files - Sync progress tracking
• summaries - Session summaries

• pattern_evidence - External validation signals (test_pass, build_ok, user_commit)
• quarantined_patterns - Provisional patterns awaiting validation
• telemetry - Context retrieval metrics (hits, misses, latency)

Add to ~/.claude/settings.json:

{
 "hooks": {
 "SessionStart": [
 {
 "command": "wsl -e bash -lc 'cd ~/.claude/claude-memory-system && bun run src/hooks/hook.ts'",
 "type": "script"
 }
 ],
 "UserPromptSubmit": [
 {
 "command": "wsl -e bash -lc 'cd ~/.claude/claude-memory-system && bun run src/hooks/hook.ts'",
 "type": "script"
 }
 ]
 }
}

This injects relevant context (commands and instincts) into your prompts automatically with a 100ms soft-block and async fallback.

bun run src/cli.ts search "docker cp" -t command

bun run src/cli.ts instincts -d deployment

bun run src/cli.ts context "deploy the whatsapp bot"

bun run src/cli.ts telemetry --hours 48

bun run src/cli.ts snapshot
bun run src/cli.ts snapshot list

bun run src/cli.ts integrity

This architecture was validated through a 6-round debate between Gemini, Codex, and Claude. Key consensus points:

1. SQLite + FTS5 - Right choice for single-user, low-write workload
2. Unified Approach - Merging 4 tools reduces complexity vs maintaining separate systems
3. Frecency Ranking - Frequency x recency is good but needs evidence-based validation
4. 100ms Soft-Block - Context injection must not block user prompts
5. Evidence Schema - External signals (tests, commits) validate patterns
6. WAL + Snapshots - Durability without complexity of replication

See ARCHITECTURE.md for the full debate synthesis.

• Database: ~/.claude-memory-system/memory.db
• Snapshots: ~/.claude-memory-system/memory-snapshot-*.db
• Source: ~/.claude/projects/*/ (Claude Code sessions)

• Bun runtime (via WSL on Windows)
• Claude Code with session logging enabled

MIT