Claude Engram

Graph intelligence engine for Claude Code — semantic search, temporal knowledge graphs, and decision lineage tracking powered by SQLite-vec.

Part of the Claude Starchart monorepo.

Capability Layers

Layer	Tools	Description
Memory	store, search, recent, delete, expand, batch	Semantic store/recall with category TTL
Graph	add_relation, get_relations, traverse_graph	Knowledge graph with typed edges
Temporal	timeline, decision_chain, record_outcome	Chronological navigation and decision lineage
Intelligence	consolidate, cache_stats, cache_clear	HDBSCAN dedup, progressive disclosure, L2 cache

Plus: hybrid search (FTS5 + vector + RRF fusion), project scoping, Unix socket IPC.

Engram vs Claude Code Native Memory

Claude Code v2.1.32+ includes built-in memory. Here's when to use each:

Capability	CC Native	Engram
Basic store/recall	Yes (automatic)	Yes
Semantic search	Basic	Advanced (hybrid FTS5+vector+RRF)
Progressive disclosure	No	Yes (index/summary/full — 10x token savings)
Knowledge graph	No	Yes (typed relations, traversal)
Decision chains	No	Yes (temporal lineage tracking)
Consolidation	No	Yes (HDBSCAN clustering dedup)
Temporal queries	No	Yes (timeline, point-in-time)
L2 semantic cache	No	Yes (query dedup, <50ms)
Cross-project	No	Yes (proj_tag scoping + global)
MCP accessible	No	Yes (any MCP client)
Category TTL	No	Yes (5min - permanent)
Setup	Zero (built-in)	Daemon + systemd

Recommendation: Use CC native memory for simple facts. Use Engram when you need search quality, decision tracking, graph relations, or cross-session intelligence.

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                        Claude Code                               │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐          │
│  │  MCP Server  │  │     CLI      │  │  Star Bridge │          │
│  └──────┬───────┘  └──────┬───────┘  └──────┬───────┘          │
└─────────┼─────────────────┼─────────────────┼───────────────────┘
          │                 │                 │
          │    JSON over Unix Socket          │
          │   /run/user/$UID/claude-engram.sock
          ▼                 ▼                 ▼
┌─────────────────────────────────────────────────────────────────┐
│                      MemoryDaemon                                │
│  ┌──────────────┐  ┌──────────────┐  ┌──────────────┐          │
│  │ Request      │  │  TTLCache    │  │  Periodic    │          │
│  │ Dispatcher   │  │  (sessions)  │  │  Cleanup     │          │
│  └──────┬───────┘  └──────────────┘  └──────────────┘          │
│         │                                                        │
│  ┌──────▼───────────────────────────────────────────┐           │
│  │              MemoryStorage                        │           │
│  │  ┌────────────┐  ┌────────────┐  ┌────────────┐  │           │
│  │  │  SQLite    │  │ sqlite-vec │  │  fastembed │  │           │
│  │  │  (WAL)     │  │ (vectors)  │  │  (MiniLM)  │  │           │
│  │  └────────────┘  └────────────┘  └────────────┘  │           │
│  │  ┌────────────┐  ┌────────────┐  ┌────────────┐  │           │
│  │  │  Graph     │  │ Semantic   │  │  FTS5      │  │           │
│  │  │  Engine    │  │ Cache (L2) │  │  (hybrid)  │  │           │
│  │  └────────────┘  └────────────┘  └────────────┘  │           │
│  └──────────────────────────────────────────────────┘           │
└─────────────────────────────────────────────────────────────────┘

Installation

# From monorepo root
pip install -e packages/engram

# Or standalone
pip install claude-engram

Systemd Service (recommended)

# Copy service file
cp platforms/linux/claude-engram.service ~/.config/systemd/user/

# Enable and start
systemctl --user enable claude-engram
systemctl --user start claude-engram

# Check status
systemctl --user status claude-engram

Usage

CLI

# Daemon control
claude-engram start              # Start daemon (foreground)
claude-engram stop               # Stop daemon
claude-engram status             # Check if running
claude-engram health             # Detailed health check

# Memory operations
claude-engram store "Decision: use PostgreSQL for persistence" --category decision
claude-engram search "database choice"
claude-engram recent --limit 5
claude-engram delete 42
claude-engram stats

MCP Tools (20)

Memory Layer

Tool	Description
search_memories	Semantic/hybrid/keyword search with progressive disclosure
store_memory	Store with category TTL and supersession tracking
recent_memories	Chronological list (no ranking)
delete_memory	Delete by ID (project-scoped)
expand_memory	Drill into a specific memory (full detail)
get_memories_batch	Retrieve multiple memories by ID

Graph Layer

Tool	Description
add_relation	Create typed edge between memories
get_relations	Query relations for a memory
traverse_graph	Multi-hop graph traversal with cycle prevention

Temporal Layer

Tool	Description
timeline	Navigate chronologically around an anchor memory
decision_chain	Follow decision lineage (resulted_in edges)
record_outcome	Link a decision to its measured outcome

Intelligence Layer

Tool	Description
consolidate_memories	HDBSCAN-based duplicate detection and cleanup
cache_stats	L2 semantic cache statistics
cache_clear	Clear L2 cache

Counters

Tool	Description
get_counters	Read named counters
set_counter	Set counter value
update_counter	Increment/decrement counter
delete_counter	Remove counter

System

Tool	Description
health_check	Daemon and database health diagnostics

Example: Claude Code MCP Config

{
  "mcpServers": {
    "engram": {
      "type": "stdio",
      "command": "python",
      "args": ["-m", "claude_engram.mcp_server"]
    }
  }
}

Memory Categories

Categories determine automatic TTL (time-to-live) for memories:

Category	TTL	Use Case
lesson	Permanent	Insights, patterns, best practices
decision	7 days	Architectural choices, trade-offs
interface	7 days	API discoveries, function signatures
file_ownership	7 days	Who owns what file
session_summary	7 days	End-of-session context
checkpoint	4 hours	Progress markers, state snapshots
query_cache	5 minutes	Temporary search results
general	Permanent	Legacy category (v1 compatibility)

TTL Tiers

PERMANENT ─────────────────────────────────────────► never expires
SESSION   ─────────────────────────────► 7 days
VOLATILE  ───────────────► 4 hours
EPHEMERAL ──► 5 min

Search Modes

Engram supports three search strategies via the search_mode parameter:

Mode	Engine	Use Case
vector (default)	Cosine similarity on MiniLM-L6-v2 embeddings	Semantic meaning
hybrid	FTS5 BM25 + vector with RRF fusion	Best of both worlds
keyword	FTS5 BM25 text matching	Exact term matching

Results are ranked by a composite score combining similarity (0.6), salience (0.2), recency decay (0.1), and access frequency (0.1).

Progressive Disclosure

Token-efficient retrieval in three levels:

Level	Content	Tokens/result
index	IDs, category, similarity, token_count only	~50
summary (default)	Index + first sentence/100 chars	~100
full	Complete memory text + metadata	~500

Use search_memories(detail_level="index") first, then expand_memory or get_memories_batch for drill-down.

Graph Relations

The knowledge graph supports six typed edge relations:

Relation	Meaning
updates	Source updates information in target
supersedes	Source replaces target (target marked inactive)
supports	Source corroborates target
contradicts	Source conflicts with target
resulted_in	Decision led to this outcome
related_to	General association

Graph traversal uses recursive CTEs with automatic cycle prevention (delimiter-wrapped path tracking).

Output Format

All MCP tools return XML-structured responses:

<memories count="3">
  <memory id="42" category="decision" similarity="87%" scope="proj:org/repo" created="2026-02-05T10:30:00">
    Decision: Use PostgreSQL for WebSocket state persistence
  </memory>
  <memory id="38" category="lesson" similarity="72%" scope="global" created="2026-02-04T15:20:00">
    Lesson: SQLite WAL mode enables concurrent reads during writes
  </memory>
</memories>

<memory_stored id="43" category="lesson" scope="proj:org/repo"></memory_stored>

<memory_deleted id="42" status="success"></memory_deleted>

Configuration

Environment Variables

Variable	Default	Description
ENGRAM_EMBED_MODEL	all-MiniLM-L6-v2	Fastembed model name
ENGRAM_EMBED_DIM	384	Embedding vector dimension
ENGRAM_EMBED_THREADS	4	ONNX threads for embedding model
ENGRAM_CLEANUP_INTERVAL	3600	Seconds between expired memory cleanup
ENGRAM_SOCKET_CHECK_INTERVAL	30	Seconds between socket health checks
ENGRAM_CHECKPOINT_LIGHT	40	Tool calls before checkpoint (light profile)
ENGRAM_CHECKPOINT_BALANCED	25	Tool calls before checkpoint (balanced)
ENGRAM_CHECKPOINT_ULTRA	15	Tool calls before checkpoint (ultra)

Socket Location

Socket path is resolved in order:

1. CLAUDE_ENGRAM_SOCKET env var
2. $XDG_RUNTIME_DIR/claude-engram.sock
3. /run/user/$UID/claude-engram.sock
4. /tmp/claude-engram.sock (fallback)

Database Location

~/.local/share/claude-engram/
├── memories.db          # SQLite database
├── memories.db-wal      # WAL file (auto-checkpointed)
├── memories.db-shm      # Shared memory
└── daemon.pid           # PID file

Database Schema

-- memories table (v10 schema)
CREATE TABLE memories (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    content TEXT NOT NULL,
    category TEXT DEFAULT 'general',
    metadata TEXT DEFAULT '{}',
    created_at TEXT NOT NULL,
    updated_at TEXT NOT NULL,
    proj_tag TEXT DEFAULT '',
    ttl_tier TEXT DEFAULT 'permanent',
    expires_at TEXT,
    content_hash TEXT UNIQUE,
    summary TEXT,
    token_count INTEGER,
    superseded_by INTEGER,
    is_active INTEGER DEFAULT 1,
    valid_from TEXT,
    valid_to TEXT,
    access_count INTEGER DEFAULT 0
);

-- Vector storage (sqlite-vec)
CREATE VIRTUAL TABLE memory_vectors USING vec0(
    id INTEGER PRIMARY KEY,
    embedding FLOAT[384]
);

-- Knowledge graph edges
CREATE TABLE memory_relations (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    source_id INTEGER NOT NULL,
    target_id INTEGER NOT NULL,
    relation_type TEXT NOT NULL,
    metadata TEXT DEFAULT '{}',
    created_at TEXT NOT NULL,
    valid_from TEXT,
    valid_to TEXT,
    invalidated_at TEXT,
    UNIQUE(source_id, target_id, relation_type)
);

-- FTS5 full-text search (external content)
CREATE VIRTUAL TABLE memory_fts USING fts5(
    content, category,
    tokenize='porter unicode61 remove_diacritics 2',
    content='memories', content_rowid='id'
);

-- Counters
CREATE TABLE counters (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    key TEXT NOT NULL,
    value INTEGER NOT NULL DEFAULT 0,
    proj_tag TEXT NOT NULL DEFAULT '',
    updated_at TEXT NOT NULL,
    UNIQUE(key, proj_tag)
);

-- Consolidation tracking
CREATE TABLE consolidation_meta (
    key TEXT PRIMARY KEY,
    value TEXT NOT NULL
);

Performance Tuning

For long-running deployments (12-24h+), configure thread limits to prevent memory growth:

# In systemd service or environment
Environment=ENGRAM_EMBED_THREADS=4
Environment=OMP_NUM_THREADS=4
Environment=HF_HUB_DISABLE_TELEMETRY=1

Why This Matters

ONNX runtime uses arena allocators that don't release memory back to the OS. Without thread limits:

• 51 threads -> 8 threads (-86%)
• 227 MB -> 151 MB (-33%)

Monitoring

# Process metrics via daemon
echo '{"action":"stats"}' | nc -U /run/user/$UID/claude-engram.sock | jq .process

# Returns:
{
  "memory_mb": 151.5,
  "threads": 2,
  "cache_entries": 0
}

# Systemd status
systemctl --user status claude-engram

Project Scoping

Memories are scoped by proj_tag for multi-project isolation:

# Auto-detected from environment
STAR_PROJECT_TAG    -> "my-project"
CLAUDE_PROJECT_DIR  -> "proj:project-name"

# Or explicit
store_memory(content="...", proj_tag="proj:my-project")

Search includes both project-specific and global memories by default.

Counters

Simple key-value counters for tracking state:

# CLI
claude-engram counter set session.tools 0
claude-engram counter update session.tools 1  # increment
claude-engram counter get session.tools

# MCP
set_counter(key="session.tools", value=0)
update_counter(key="session.tools", delta=1)
get_counters()

Troubleshooting

Daemon won't start

# Check if already running
claude-engram status

# Check socket
ls -la /run/user/$UID/claude-engram.sock

# Check logs
journalctl --user -u claude-engram -f

Socket deleted (orphan state)

The daemon auto-recovers from orphan sockets (checked every 30s). If manual recovery needed:

systemctl --user restart claude-engram

High memory usage

Check thread count and configure limits:

# Current threads
cat /proc/$(pgrep -f claude_engram.daemon)/status | grep Threads

# Apply limits
export ENGRAM_EMBED_THREADS=4
systemctl --user restart claude-engram

License

MIT