engram-mem 0.5.17

Memory traces for AI agents - Think like human

engram

Memory traces for AI agents — Think like humans.

Dual-memory AI system combining episodic (vector) + semantic (graph) memory with LLM reasoning. Enterprise-ready with multi-tenancy, auth, caching, observability, and Docker deployment. Exposes CLI, MCP, HTTP API (/api/v1/), and WebSocket API (/ws) for real-time agent communication.

---

Features

Core Memory

• Episodic Memory — ChromaDB vector store, semantic similarity search, Ebbinghaus decay, activation-based scoring, topic-key upsert
• Semantic Graph — NetworkX MultiDiGraph, typed entities and relationships, SQLite (default) or PostgreSQL backend, weighted edges
• Reasoning Engine — LLM synthesis (Gemini via litellm), dual-memory context fusion, constitution-guarded prompts
• Recall Pipeline — Query decision, temporal+pronoun entity resolution, parallel multi-source search, dedup, composite scoring
• Auto Memory — Detect and persist save-worthy messages automatically, poisoning guard for injection prevention
• Feedback Loop — Confidence scoring (+0.15/−0.2), importance adjustment, auto-delete on 3x negative feedback
• Graph Visualization — Interactive entity relationship explorer with dark theme, search, click-to-inspect (vis-network)

Intelligence Layer

• Temporal Resolution — 28 Vietnamese+English date patterns resolve "hôm nay/yesterday" → ISO dates before storing
• Pronoun Resolution — "anh ấy/he/she" → named entity from graph context, LLM-based fallback
• Fusion Formatter — Group recall results by type [preference]/[fact]/[lesson] for structured LLM context
• Memory Consolidation — Jaccard clustering + LLM summarization reduces redundancy

Enterprise

• Multi-Surface — CLI (Typer), MCP Server (stdio), HTTP API (FastAPI), WebSocket, Web UI
• Authentication — JWT + API keys with RBAC (ADMIN, AGENT, READER), optional, disabled by default
• Multi-Tenancy — Isolated per-tenant stores, contextvar propagation, row-level PostgreSQL isolation
• Caching — Redis-backed result caching with per-endpoint TTLs
• Rate Limiting — Sliding-window per-tenant limits, fail_open option
• Audit Trail — Structured before/after JSONL log for every episodic mutation
• Resource Tiers — 4-tier LLM degradation (FULL → STANDARD → BASIC → READONLY), 60s auto-recovery
• Data Constitution — 3-law LLM governance (namespace isolation, no fabrication, audit rights), SHA-256 tamper detection
• Consolidation Scheduler — Asyncio background tasks (cleanup daily, consolidate 6h, decay daily), tier-aware
• Key Rotation — Failover/round-robin for embedding API keys (GEMINI_API_KEY + GEMINI_API_KEY_FALLBACK)
• Observability — OpenTelemetry + JSONL audit logging (optional)
• Deployment — Docker Compose, Kubernetes-ready, health checks
• Backup/Restore — Memory snapshots, point-in-time recovery
• Benchmark Suite — p50/p95/p99 latency measurements for all endpoints

---

Architecture

flowchart TD
    CLI["CLI (Typer)"]
    MCP["MCP (stdio)"]
    HTTP["HTTP API /api/v1/"]
    WS["WebSocket /ws"]

    CLI & MCP & HTTP & WS --> Auth["Auth Middleware\n(JWT + RBAC, optional)"]
    Auth --> Tenant["TenantContext (ContextVar)"]
    Tenant --> Recall["Recall Pipeline\n(decision → resolve → search → feedback)"]
    Recall --> Episodic["EpisodicStore\n(ChromaDB)"]
    Recall --> Semantic["SemanticGraph\n(NetworkX + SQLite/PG)"]
    Episodic & Semantic --> Reasoning["Reasoning Engine\n(Gemini via litellm)"]
    Episodic --> Cache["Redis Cache (optional)"]
    WS --> EventBus["Event Bus\n(push events)"]

---

Quick Start

# Install from source
git clone https://github.com/docaohieu2808/Engram-Mem.git
cd engram
pip install -e .

# Optional: dev dependencies
pip install -e ".[dev]"

# Optional: OpenTelemetry support
pip install -e ".[telemetry]"

Requirements: Python 3.11+

Optional: GEMINI_API_KEY for LLM reasoning and embeddings. Basic storage works without it.

CLI Usage

export GEMINI_API_KEY="your-key"

# Store a memory
engram remember "Deployed v2.1 to production at 14:00 - caused 503 spike"

# Retrieve similar memories
engram recall "production incidents"

# Reason across all memory
engram think "What deployment issues have we had?"

# Add knowledge graph entities
engram add node "PostgreSQL" --type Technology
engram add edge "Service:API" "Technology:PostgreSQL" --relation uses

# Start HTTP API server
engram serve

Python SDK

from engram import EngramClient

async with EngramClient(namespace="my-agent") as client:
    # Auto-recalls relevant context before LLM call,
    # auto-extracts facts from response in the background
    response = await client.chat([
        {"role": "user", "content": "Deploy to production"}
    ])

# Explicit operations
await client.remember("No deploys on Fridays", memory_type="decision", priority=8)
results = await client.recall("deployment policy", limit=5)
answer = await client.think("What are our deployment rules?")

HTTP API Usage

# Store memory
curl -X POST http://localhost:8765/api/v1/remember \
  -H "Content-Type: application/json" \
  -d '{"content": "Deployed v1.0 to production", "memory_type": "fact", "priority": 8}'

# Search
curl "http://localhost:8765/api/v1/recall?query=deployment&limit=5"

# Reason
curl -X POST http://localhost:8765/api/v1/think \
  -H "Content-Type: application/json" \
  -d '{"question": "What deployment issues have we had?"}'

---

Configuration

Config file: ~/.engram/config.yaml — Priority: CLI flags > env vars > YAML > defaults

episodic:
  provider: chromadb
  path: ~/.engram/episodic
  namespace: default

embedding:
  provider: gemini
  model: gemini-embedding-001
  # Key rotation: failover (default) or round-robin
  key_strategy: failover  # GEMINI_API_KEY + GEMINI_API_KEY_FALLBACK

semantic:
  provider: sqlite          # or postgresql
  path: ~/.engram/semantic.db
  # For PostgreSQL:
  # dsn: postgresql://user:pass@localhost/engram
  # pool_min: 5
  # pool_max: 20

llm:
  provider: gemini
  model: gemini/gemini-2.0-flash
  api_key: ${GEMINI_API_KEY}

auth:
  enabled: false            # Set to true for production
  jwt_secret: "use-32+-chars-or-${ENV_VAR}"

cache:
  enabled: false            # Set to true with Redis
  redis_url: redis://localhost:6379/0

rate_limit:
  enabled: false
  requests_per_minute: 100
  fail_open: true           # Allow through on Redis failure

audit:
  enabled: false
  path: ~/.engram/audit.jsonl

telemetry:
  enabled: false            # Requires telemetry extra
  otlp_endpoint: http://localhost:4317

recall:
  enabled: true
  decision_skip_trivial: true
  entity_resolution_enabled: true
  parallel_search_enabled: true
  feedback_enabled: true
  auto_consolidate_threshold: 20
  retrieval_audit_enabled: true

ingestion:
  auto_memory_enabled: true
  guard_enabled: true

feedback:
  confidence_positive_delta: 0.15
  confidence_negative_delta: 0.2
  auto_delete_threshold: 3

resolution:
  temporal_enabled: true
  pronoun_enabled: true

fusion:
  formatter_enabled: true

graph:
  visualization_enabled: true

---

Environment Variables

Variable	Purpose
GEMINI_API_KEY	LLM + embeddings (primary key)
GEMINI_API_KEY_FALLBACK	Secondary key for key rotation
GEMINI_KEY_STRATEGY	failover (default) or round-robin
ENGRAM_AUTH_ENABLED	Enable auth (true/false)
ENGRAM_AUTH_JWT_SECRET	JWT signing key (32+ chars)
ENGRAM_SEMANTIC_PROVIDER	sqlite or postgresql
ENGRAM_SEMANTIC_DSN	PostgreSQL connection string
ENGRAM_CACHE_ENABLED	Enable Redis caching
ENGRAM_CACHE_REDIS_URL	Redis URL
ENGRAM_RATE_LIMIT_ENABLED	Enable rate limiting
ENGRAM_RATE_LIMIT_REQUESTS_PER_MINUTE	Default 60
ENGRAM_AUDIT_ENABLED	Enable audit logs
ENGRAM_TELEMETRY_ENABLED	Enable OpenTelemetry
ENGRAM_RECALL_ENABLED	Enable recall pipeline
ENGRAM_INGESTION_AUTO_MEMORY_ENABLED	Detect save-worthy messages
ENGRAM_INGESTION_GUARD_ENABLED	Block prompt injection
ENGRAM_FEEDBACK_ENABLED	Enable feedback loop
ENGRAM_RESOLUTION_TEMPORAL_ENABLED	Temporal entity resolution
ENGRAM_RESOLUTION_PRONOUN_ENABLED	Pronoun resolution via LLM
ENGRAM_FUSION_FORMATTER_ENABLED	Format recall by type
ENGRAM_GRAPH_VISUALIZATION_ENABLED	Enable graph UI at /graph

Note: ${VARIABLE} syntax in YAML is expanded at load time.

---

API Reference

Start server: engram serve [--host 0.0.0.0] [--port 8765]

All endpoints at /api/v1/; legacy routes redirect. Auth disabled by default.

Method	Endpoint	Purpose
GET	/health	Liveness check (always available)
GET	/health/ready	Readiness probe
POST	/api/v1/remember	Store episodic memory
GET	/api/v1/recall	Search memories (?query=X&limit=5&offset=0)
POST	/api/v1/think	LLM reasoning across episodic + semantic
GET	/api/v1/query	Graph search (?keyword=X&node_type=Y&related_to=Z)
POST	/api/v1/ingest	Extract entities + store memories from messages
POST	/api/v1/feedback	Record positive/negative feedback on a memory
GET	/api/v1/graph/data	Graph data JSON for visualization
GET	/graph	Interactive graph visualization UI
POST	/api/v1/cleanup	Delete expired memories (admin only)
POST	/api/v1/summarize	LLM synthesis of recent memories (admin only)
POST	/api/v1/auth/token	Issue JWT (admin_secret required)
POST	/api/v1/backup	Export all memory to JSON
POST	/api/v1/restore	Import backup snapshot

Responses: All wrapped in {data, meta}. Errors: {error: {code, message}, meta}.

Auth: When enabled, use Authorization: Bearer <JWT> or X-API-Key: <key>.

---

WebSocket API

Real-time bidirectional communication. Connect via ws://host:8765/ws?token=JWT (token optional when auth disabled).

Commands (client to server):

Command	Payload
remember	{"content": "...", "priority": 7}
recall	{"query": "...", "limit": 5}
think	{"question": "..."}
feedback	{"memory_id": "abc123", "feedback": "positive"}
query	{"keyword": "PostgreSQL"}
ingest	{"messages": [...]}
status	{}

Example:

{"id": "corr-1", "type": "remember", "payload": {"content": "User prefers dark mode", "priority": 7}}
{"id": "corr-1", "type": "response", "status": "ok", "data": {"id": "mem-xyz"}}

Push Events (server to all agents in same tenant):

{"type": "event", "event": "memory_created", "tenant_id": "default", "data": {"id": "mem-xyz"}}

Events: memory_created, memory_updated, memory_deleted, feedback_recorded

---

CLI Reference

Memory Operations

# Store with options
engram remember <content> [--type fact|decision|preference|todo|error|context|workflow]
                          [--priority 1-10] [--tags tag1,tag2] [--expires 2h|1d|7d]

# Search
engram recall <query> [--limit 5] [--type <type>] [--tags tag1,tag2] [--namespace <ns>]
              [--resolve-entities] [--resolve-temporal]

# Reason
engram think <question>
engram summarize [--count 20] [--save]

Semantic Graph

engram add node <name> --type <NodeType>
engram add edge <from_key> <to_key> --relation <relation>
engram remove node <key>
engram query [<keyword>] [--type <NodeType>] [--related-to <name>] [--format table|json]
engram graph [--search <keyword>]       # Open interactive graph visualization

Intelligence & Pipeline

engram resolve <query> [--context "..."]      # Entity/temporal resolution
engram feedback <id> [--positive|--negative]  # Record feedback on memory
engram audit [--last N]                        # View retrieval audit log
engram benchmark --questions file.json         # Run accuracy benchmarks

System

engram status                          # Memory + graph stats
engram serve [--host 127.0.0.1] [--port 8765]
engram watch [--daemon]                # Watch inbox + start scheduler
engram resource-status                 # Show resource tier (FULL/STANDARD/BASIC/READONLY)
engram constitution-status             # Show 3-law governance + SHA-256 hash
engram scheduler-status                # Show background task schedule + state

Maintenance

engram cleanup                         # Delete expired memories
engram ingest <file.json> [--dry-run]  # Ingest chat JSON
engram backup                          # Export memory snapshot
engram restore <file>                  # Import snapshot
engram config show / get <key> / set <key> <value>

---

MCP Integration

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "engram": {
      "command": "/path/to/.venv/bin/engram-mcp",
      "env": {
        "GEMINI_API_KEY": "your-key"
      }
    }
  }
}

Available MCP Tools:

Tool	Description
engram_remember	Store memory with type, priority, tags, namespace
engram_recall	Search episodic memories (compact format by default)
engram_resolve	Entity/temporal resolution for query context
engram_feedback	Record positive/negative feedback on memories
engram_think	Reason across episodic + semantic memory via LLM
engram_summarize	Summarize recent N memories via LLM
engram_cleanup	Delete all expired memories
engram_status	Show memory statistics
engram_get_memory	Retrieve full memory content by ID or prefix
engram_timeline	Chronological context around a memory
engram_add_entity	Add entity node to knowledge graph
engram_add_relation	Add relationship edge between entities
engram_query_graph	Query knowledge graph
engram_get_graph_data	Retrieve graph data for visualization
engram_ingest	Dual ingest: extract entities + store memories
engram_session_start	Begin new conversation session
engram_session_end	End active session with optional summary
engram_session_summary	Get summary of completed session
engram_session_context	Retrieve memories from active session

Claude Code Hooks: Engram ships Stop + SessionEnd hooks that capture conversation turns automatically — no SDK integration needed.

---

Embeddings

Only one embedding model is supported:

Mode	Model	Dimensions	Requires
Gemini (default)	gemini-embedding-001	3072	GEMINI_API_KEY
Fallback	all-MiniLM-L6-v2 (ChromaDB default)	384	nothing

Embedding dimensions must remain consistent within a collection. If you switch providers, reinitialize the episodic store or create a new collection.

Key rotation: Set GEMINI_API_KEY_FALLBACK and GEMINI_KEY_STRATEGY=round-robin to distribute quota across multiple API keys.

---

Benchmarks

Run python tests/benchmark_performance.py --host 127.0.0.1 --port 8765 against a running server.

Sample results (local, SQLite backend, no Redis):

Operation	p50	p95	p99	Notes
GET /health	0.8ms	1.2ms	2.1ms	Always fast
POST /remember	415ms	680ms	920ms	Embedding API bound
GET /recall	1.3ms	3.2ms	5.8ms	Vector search
POST /think	5.6s	7.2s	9.4s	LLM bound

Flags: --quick for a faster subset, --concurrency N for load testing.

---

Docker

Quick Start

docker build -t engram:latest .
docker run -e GEMINI_API_KEY="your-key" -p 8765:8765 engram:latest

Production (PostgreSQL + Redis)

# Production env vars
ENGRAM_AUTH_ENABLED=true
ENGRAM_AUTH_JWT_SECRET=$(openssl rand -hex 32)
ENGRAM_SEMANTIC_PROVIDER=postgresql
ENGRAM_SEMANTIC_DSN=postgresql://user:pass@postgres:5432/engram
ENGRAM_CACHE_ENABLED=true
ENGRAM_CACHE_REDIS_URL=redis://redis:6379/0
ENGRAM_AUDIT_ENABLED=true
GEMINI_API_KEY=your-key
GEMINI_API_KEY_FALLBACK=your-fallback-key
GEMINI_KEY_STRATEGY=round-robin

See deployment-guide.md for Docker Compose with PostgreSQL, Redis, and OpenTelemetry.

---

Test Coverage

# Run all tests
pytest tests/ -v

# With coverage
pytest tests/ --cov=src/engram --cov-report=html

# Specific suites
pytest tests/ -k "recall or resolution or feedback" -v
pytest tests/ -k "websocket" -v

• 894+ tests across all modules
• 61%+ code coverage (core features 80%+)
• CI/CD: GitHub Actions runs full suite on every PR and commit

---

Documentation

• Project Overview & PDR — Features, requirements, config reference
• System Architecture — Design, data flow, deployment patterns
• Code Standards — Conventions, patterns, best practices
• Deployment Guide — Docker, Kubernetes, environment variables, auth setup
• Codebase Summary — Module inventory, metrics
• Project Roadmap — Completed phases and future work
• Changelog — Full version history

---

Contributing

Pull requests welcome. See CONTRIBUTING.md.

Before submitting:

• Run ruff check src/ for linting
• Run pytest tests/ — do not ignore failing tests
• Do not commit .env files or API keys

---

License

MIT — Copyright (c) Do Cao Hieu