Engram MCP Server

Durable memory for AI agents — temporal knowledge graph, hybrid retrieval, SQLite or PostgreSQL.

License

java-ai-memory.dev · Source code · JamJet docs · Discord

Engram is a durable memory layer for AI agents. It extracts facts from conversations, stores them in a temporal knowledge graph, and retrieves them with hybrid semantic + keyword search — backed by a single SQLite file or a PostgreSQL database.

This repo hosts the Glama registry listing. Source code lives in the main JamJet repo.

Quickstart — 30 seconds

# Docker — uses local Ollama by default
docker run --rm -i \
  -v engram-data:/data \
  ghcr.io/jamjet-labs/engram-server:0.5.0

Or install from crates.io:

cargo install jamjet-engram-server
engram serve

Claude Desktop configuration

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "engram": {
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-v", "engram-data:/data",
        "ghcr.io/jamjet-labs/engram-server:0.5.0"
      ]
    }
  }
}

After restart, 11 MCP tools are available to the model.

MCP Tools (11)

Memory tools (7)

Tool	Description
`memory_add`	Extract and store facts from conversation messages using LLM-powered fact extraction. Side effects: calls the configured LLM to parse facts, then writes them to the knowledge graph. Returns extracted fact IDs. Requires `messages` array and `user_id`.
`memory_recall`	Semantic search over stored facts using vector similarity. Read-only, no side effects. Returns ranked facts matching the query, scoped by `user_id` and optional `org_id`. Use this to retrieve relevant context before generating a response.
`memory_context`	Assemble a token-budgeted context block for LLM prompts with tier-aware fact selection. Read-only. Returns a formatted string of the most relevant facts, capped at the specified token budget. Use this instead of memory_recall when you need a ready-to-use prompt snippet.
`memory_search`	Keyword search over facts using full-text search (SQLite FTS5 / Postgres). Read-only, no side effects. Returns facts matching exact keywords. Use this when you need precise term matching rather than semantic similarity from memory_recall.
`memory_forget`	Soft-delete a fact by ID with an optional reason. Side effect: marks the fact as deleted in the knowledge graph (does not physically remove it). Irreversible via this tool. Use when a user asks to remove specific information.
`memory_stats`	Get aggregate statistics: total facts, valid (non-deleted) facts, entity count, and relationship count. Read-only, no side effects. Use this to understand the size and health of the memory store.
`memory_consolidate`	Run a maintenance cycle over the knowledge graph — decay stale facts, promote high-confidence ones, deduplicate near-duplicates, and summarize clusters. Side effects: modifies fact scores and may merge or archive facts. Run periodically to keep memory accurate.

Message store tools (4)

Tool	Description
`messages_save`	Save chat messages for a conversation by ID. Side effects: writes messages to the store and optionally triggers fact extraction (controlled by `--extract-on-save`). Use this to persist full conversation history alongside extracted facts.
`messages_get`	Retrieve all messages for a conversation by ID. Read-only, no side effects. Returns the ordered message array. Use this to replay or inspect a past conversation.
`messages_list`	List all conversation IDs in the message store. Read-only, no side effects. Returns an array of conversation ID strings. Use this to discover what conversations are stored before retrieving with messages_get.
`messages_delete`	Delete all messages for a conversation by ID. Side effect: permanently removes the conversation's messages from the store. Irreversible. Does not affect extracted facts — use memory_forget for that.

All memory tools are scoped by (org_id, user_id, session_id) — org is the coarsest, session the finest.

LLM Providers

Provider-agnostic. One binary, set ENGRAM_LLM_PROVIDER=... and go:

Provider	Env value	Notes
Ollama	`ollama` (default)	Local, free, no API keys
OpenAI-compatible	`openai-compatible`	OpenAI, Azure, Groq, Together, Mistral, DeepSeek, vLLM, LM Studio, ...
Anthropic	`anthropic`	Claude via Messages API
Google	`google`	Gemini via generateContent
Shell command	`command`	Pipe to any external script
Mock	`mock`	Deterministic, for tests only

# Example: use Groq instead of Ollama
docker run --rm -i \
  -e ENGRAM_LLM_PROVIDER=openai-compatible \
  -e ENGRAM_OPENAI_BASE_URL=https://api.groq.com/openai/v1 \
  -e OPENAI_API_KEY=gsk_... \
  -v engram-data:/data \
  ghcr.io/jamjet-labs/engram-server:0.5.0

Why Engram?

Problem	Engram's answer
Every agent memory library is Python-first	Rust core with native Python, Java, and MCP clients
Needs Postgres + Qdrant + Neo4j just to try	Single SQLite file (zero infra) or Postgres when you need it
Conversation history is not knowledge memory	Fact extraction pipeline — structured facts from messages
Old facts drift and contradict	Conflict detection + consolidation — decay, promote, dedup, summarize
Memory recall is either semantic OR keyword	Hybrid retrieval — vector search + FTS5 in one query
MCP support is an afterthought	MCP-native — 11 tools exposed by a single binary
Can't isolate memory per user or tenant	First-class scopes — org / user / session built into every query

Client SDKs

Language	Package	Install
Python	`jamjet` (includes `EngramClient`)	`pip install jamjet`
Java	`dev.jamjet:jamjet-sdk` (includes `EngramClient`)	Maven Central
Spring Boot	`dev.jamjet:engram-spring-boot-starter`	Maven Central
Rust	`jamjet-engram` (embed directly)	`cargo add jamjet-engram`

JamJet — the full agent-native runtime (parent project)
java-ai-memory.dev — comparison with Mem0, Zep, LangChain4j, Spring AI, and others
Full Engram docs

License

Apache 2.0 — see LICENSE.

Engram MCP Server