Engram Server

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.

It ships in two shapes:

• jamjet-engram — a Rust library you embed in your own application.
• jamjet-engram-server (this crate) — a standalone binary that speaks MCP over stdio and REST over HTTP, so Claude Desktop, Cursor, and any HTTP client can use it with no code.

Engram is provider-agnostic. Five LLM backends are wired in out of the box, and a sixth — command — lets you shell out to any external script, so you can plug in a provider Engram does not ship natively without touching Rust code:

Pick one with ENGRAM_LLM_PROVIDER=… — the same binary handles all of them.

State of the project, April 2026. Engram is new — v0.3.2, small community, no public LongMemEval / DMR numbers yet. The architecture below works, the tests pass, the Docker image runs. If you need production-scale memory today, Mem0 Cloud and Zep Cloud are more mature. If you need a tryable, self-hostable, single-binary memory layer that doesn't require Python, Postgres, Qdrant, or Neo4j, Engram is built for you.

Why Engram?

Quickstart — 30 seconds

Engram speaks to four LLM providers out of the box: Ollama (default — free, local), OpenAI, Anthropic Claude, and Google Gemini. Pick one.

Option A — Ollama (free, local, zero API keys)

Requirements: Ollama with llama3.2 and nomic-embed-text pulled:

ollama pull llama3.2ollama pull nomic-embed-text

Docker (easiest):

docker run --rm -i \  -v engram-data:/data \  ghcr.io/jamjet-labs/engram-server:0.3.2

The container defaults to Ollama at host.docker.internal:11434 — works out of the box on Docker Desktop for Mac and Windows. Linux users need --add-host=host.docker.internal:host-gateway.

Option B — Anthropic Claude (hosted, highest quality)

docker run --rm -i \  -e ENGRAM_LLM_PROVIDER=anthropic \  -e ANTHROPIC_API_KEY=sk-ant-... \  -v engram-data:/data \  ghcr.io/jamjet-labs/engram-server:0.3.2

Defaults to claude-haiku-4-5-20251001. Override with -e ENGRAM_ANTHROPIC_MODEL=claude-sonnet-4-6.

Note: Anthropic has no native JSON mode, so Engram parses JSON from text responses. The llm_util::extract_json_payload helper strips markdown fences that Claude occasionally emits.

Option C — Google Gemini

docker run --rm -i \  -e ENGRAM_LLM_PROVIDER=google \  -e GOOGLE_API_KEY=AIza... \  -v engram-data:/data \  ghcr.io/jamjet-labs/engram-server:0.3.2

Defaults to gemini-flash-latest. Override with -e ENGRAM_GOOGLE_MODEL=gemini-2.5-flash.

Option D — Any OpenAI-compatible endpoint

docker run --rm -i \  -e ENGRAM_LLM_PROVIDER=openai-compatible \  -e OPENAI_API_KEY=sk-... \  -v engram-data:/data \  ghcr.io/jamjet-labs/engram-server:0.3.2

Defaults to OpenAI itself (https://api.openai.com/v1, gpt-4o-mini). Change one env var to point at any of these providers without recompiling:

openai is accepted as a backwards-compatible alias for openai-compatible.

Option E — command (shell out to anything)

For providers Engram does not ship natively — an internal model behind a custom RPC, a raw SOAP endpoint, a local inference binary, a quick wrapper over a Python SDK, or tomorrow's new provider — use the command backend. Engram will spawn your script per extraction call, pipe a JSON request to its stdin, and read a JSON response from its stdout.

The contract. Your script reads one JSON object from stdin:

{"system": "<extraction prompt>", "user": "<conversation text>", "structured": true}

Then writes one JSON value to stdout. Either a bare value:

{"facts": [{"text": "...", "entities": [...], "confidence": 0.95, "category": "..."}]}

Or an envelope (useful for surfacing errors):

{"content": {"facts": [...]}}{"error": "rate limited, try again"}

Exit 0 on success, non-zero (with stderr) on failure.

Example Python wrapper (~15 lines):

#!/usr/bin/env python3# my-llm.py — wraps any Python SDK for Engram.import json, sysfrom my_llm_sdk import chat  # your SDK
req = json.loads(sys.stdin.read())resp = chat(    system=req["system"],    user=req["user"],    json_mode=req.get("structured", False),)sys.stdout.write(json.dumps(resp))

Point Engram at it:

docker run --rm -i \  -e ENGRAM_LLM_PROVIDER=command \  -e ENGRAM_LLM_COMMAND="python /scripts/my-llm.py" \  -v $(pwd)/scripts:/scripts \  -v engram-data:/data \  ghcr.io/jamjet-labs/engram-server:0.3.2

Timeout is 120 seconds by default — override with ENGRAM_LLM_COMMAND_TIMEOUT=30.

Security: the command provider runs arbitrary commands as the Engram process user. Never use it in a multi-tenant deployment where untrusted users can set ENGRAM_LLM_COMMAND. It is a local and single-tenant feature.

Option F — cargo install (all providers same binary)

cargo install jamjet-engram-serverengram serve                                            # ollama (default)ENGRAM_LLM_PROVIDER=anthropic ANTHROPIC_API_KEY=... engram serveENGRAM_LLM_PROVIDER=command ENGRAM_LLM_COMMAND="python /path/to/wrapper.py" engram serve

Option G — REST mode for testing

engram serve --mode rest --port 9090

curl -X POST http://localhost:9090/v1/memory \  -H 'content-type: application/json' \  -d '{    "user_id": "alice",    "messages": [      {"role": "user", "content": "I am allergic to peanuts and I love sourdough."}    ]  }'
curl 'http://localhost:9090/v1/memory/recall?q=food%20allergies&user_id=alice'

No server. No config. No Python sidecar. One binary.

MCP client configuration

Claude Desktop (~/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:0.3.2"      ]    }  }}

Cursor / any MCP-aware IDE

Point it at the same docker run command, or at a locally-installed engram serve binary. After restart, eleven tools are available to the model (seven memory_* + four messages_*).

MCP tools

Memory tools (7)

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

Message store tools (4)

REST API

Thirteen endpoints, rooted at /v1/memory. Full OpenAPI-style surface:

Configuration

All settings flow through CLI flags or environment variables. Env vars are the recommended way to configure the Docker image.

Core

Ollama

OpenAI-compatible (OpenAI, Groq, Together, Mistral, DeepSeek, Azure, vLLM, …)

Anthropic

Google Gemini

Command (shell-out)

The mock backend returns empty facts and deterministic byte-cycled vectors. It exists for tests and CI. The server prints a clear warning on startup if it detects mock mode in use. API keys and command contents are never echoed to stdout or logs — the server only describes provider model at base_url or command \<first 60 chars>``.

Database Backend

Engram supports SQLite (default) and PostgreSQL backends. The backend is selected by the --db URL:

# SQLite (default)engram serve --mode rest --db engram.db
# PostgreSQLengram serve --mode rest --db postgres://user:pass@localhost:5432/engram

When using PostgreSQL, tables are created automatically on first startup.

Chat Message Store

Engram can store raw chat messages alongside extracted facts. This enables Spring AI ChatMemoryRepository integration.

REST Endpoints:

MCP Tools: messages_save, messages_get, messages_list, messages_delete

Fact extraction on message save is controlled by --extract-on-save (default: true).

Embedding the library

If you don't want a separate process, depend on jamjet-engram directly:

[dependencies]jamjet-engram = "0.3.2"

use engram::{Memory, OllamaEmbeddingProvider, OllamaLlmClient, Scope, ExtractionConfig, Message};
let embedding = Box::new(OllamaEmbeddingProvider::new());let memory = Memory::open("sqlite:engram.db?mode=rwc", embedding).await?;
let scope = Scope::user("default", "alice");let messages = vec![Message {    role: "user".into(),    content: "I am allergic to peanuts.".into(),}];
let llm = Box::new(OllamaLlmClient::new());let fact_ids = memory    .add_messages(&messages, scope.clone(), llm, ExtractionConfig::default())    .await?;
// Later, recalllet facts = memory.recall(&engram::memory::RecallQuery {    query: "food allergies".into(),    scope: Some(scope),    max_results: 5,    as_of: None,    min_score: None,}).await?;

Other languages

Engram is part of the JamJet ecosystem. These clients speak to engram-server over REST:

See java-ai-memory.dev for how Engram compares to LangChain4j ChatMemory, Spring AI ChatMemory, Koog, Embabel, Google ADK Memory Bank, the Mem0 community wrapper, and Zep — with honest notes on where Engram fits and where the alternatives are more mature.

Architecture

┌──────────────────────────────────────────────────────────┐│           Clients: Claude Desktop, Cursor, JamJet,        ││           Python SDK, Java SDK, Spring Boot, cURL         │├─────────────────┬────────────────────────────────────────┤│   MCP (stdio)   │           REST (HTTP + JSON)            │├─────────────────┴────────────────────────────────────────┤│                     engram-server                         ││                 (jamjet-engram-server)                    │├──────────────────────────────────────────────────────────┤│                       engram (lib)                        ││   Extraction pipeline  │  Hybrid retrieval  │  Scopes     ││   Conflict detection   │  Consolidation     │  Context    │├──────────────────────────────────────────────────────────┤│    LlmClient trait        │    EmbeddingProvider trait    ││    Ollama · Mock          │    Ollama · Mock              │├──────────────────────────────────────────────────────────┤│                       Storage                             ││   SQLite or PostgreSQL (facts, entities, FTS, vectors)   │└──────────────────────────────────────────────────────────┘

What Engram is not

• Not just a chat-history store. Engram now includes a message store (see Chat Message Store), but its primary value is the fact extraction and knowledge graph layer on top. If all you need is the last N messages of a conversation, a framework's built-in window may be simpler.
• Not a state checkpointer. If you need to snapshot agent execution state for resume and replay, that's what LangGraph, Koog persistence, or JamJet's own durable runtime does. Pair them with Engram — they solve different problems.
• Not a managed service. No hosted plane, no auth layer beyond scopes, no SLA. Bring your own process manager.
• Not benchmarked yet. LongMemEval and DMR numbers are on the roadmap. Until they exist, treat comparative claims with the skepticism they deserve.

Roadmap

License

Apache 2.0 — see LICENSE.