Engrama

Graph-based long-term memory framework for AI agents.

Engrama gives any AI agent persistent, structured memory backed by a knowledge graph. Instead of flat key-value stores or opaque vector databases, Engrama stores entities, observations, and relationships — and lets agents traverse that graph to reason about their accumulated knowledge.

Two backends are first-class:

• SQLite + sqlite-vec (default since 0.9) — single file, zero external services, pip install engrama and you're running.
• Neo4j 5.26 LTS (opt-in) — for multi-process production setups, large-scale vector search, or teams that already use Cypher.

The data model is identical on both. See docs/backends.md for a full decision guide; the rest of this README assumes the SQLite default.

Since 0.13.0, every node and relation is owned by an (org_id, user_id) identity and reads are fail-closed: a missing or partial scope matches nothing rather than falling back to "see all". A single-process install runs as one stable standalone identity and needs no configuration; a multi-tenant deployment supplies the identity per request from an authenticating gateway. Each identity can permanently erase its own memory through the engrama_gdpr_forget tool (GDPR right-to-erasure). See docs/security.md.

Inspired by Karpathy's second-brain concept, but built for agents instead of humans — and with graphs instead of wikis.

Why graphs?

Prerequisites

You need two things to run on the default SQLite backend. Docker is not required unless you opt into Neo4j.

Windows users: after installing Python, make sure "Add Python to PATH" is checked. After installing uv, you may need to restart your terminal.

Optional:

• Obsidian — for vault sync features.
• A local embedder for semantic search.
• Docker Desktop — only if you opt into the Neo4j backend.

Quick start (SQLite, zero-dep)

Step 1: Install

From PyPI (recommended):

pip install engrama          # or: uv add engrama

Or from source, for development:

git clone https://github.com/scops/engrama
cd engrama
uv sync

The commands below assume a PyPI install (engrama ...). From a source checkout, prefix each one with uv run (uv run engrama ...).

Step 2: Initialise the schema

engrama init --profile developer

Step 3: Verify

engrama verify

Step 4: Use it

A) From Python:

from engrama import Engrama

with Engrama() as eng:
    eng.remember("Technology", "FastAPI", "High-performance async framework")
    eng.associate("MyProject", "Project", "USES", "FastAPI", "Technology")
    results = eng.search("microservices")

B) From the command line:

engrama search "FastAPI"
engrama reflect

Quick start (Neo4j, opt-in)

If you need multi-process writes, very large vector indexes, or an existing Cypher toolchain, install with the Neo4j extra:

pip install "engrama[neo4j]"     # or, from source: uv sync --extra neo4j

Configure your credentials by copying .env.example to .env and setting GRAPH_BACKEND=neo4j. Start Neo4j with docker compose up -d, and then initialize the schema:

engrama init --profile developer
engrama verify

Hosted (beta)

Engrama is built to be self-hosted: pip install, and it runs on your machine with your data in a local file. If you would rather not operate and secure a remote instance yourself, engrama.sinensia.com is a managed deployment of this same server — EU-hosted, OAuth 2.1 in front of /mcp, per-identity isolation and GDPR erasure already wired, with setup guides for Claude, ChatGPT, Gemini, Cursor, VS Code, Windsurf, Zed and Codex.

It is invitation-only during the beta: request access. The open-source server in this repository is not a cut-down edition of it — same tools, same graph model, same code.

Security considerations

Engrama stores everything an agent learns, so treat the memory graph as sensitive data. The full policy lives in docs/security.md; the essentials:

• Data residency. On the default SQLite backend all data lives in a single local file (~/.engrama/engrama.db by default) — nothing leaves your machine. It is plain SQLite: keep it off shared filesystems, back it up, and rely on filesystem permissions for at-rest protection. On the Neo4j backend, data resides wherever you host Neo4j; you own that deployment and its region.
• Neo4j authentication. Supply credentials through .env / environment variables (NEO4J_URI, NEO4J_USERNAME, NEO4J_PASSWORD) — never commit them. The shipped docker-compose.yml is for local dev; change the default password and enable TLS before any networked use.
• Embedding providers. Endpoints reached via OPENAI_BASE_URL should use HTTPS unless they are on localhost or a trusted network. With EMBEDDING_PROVIDER=none no text is sent anywhere; search degrades to fulltext-only.
• Tenant isolation. Since 0.13.0 every node and relation is owned by an (org_id, user_id) identity and reads are fail-closed. A single install runs as one stable standalone identity; a multi-tenant deployment must inject the identity per request from an authenticating gateway (set ENGRAMA_REQUIRE_IDENTITY=1 to fail closed on missing headers). See docs/security.md.
• Right to erasure. Each identity can permanently erase its own memory via the engrama_gdpr_forget tool (GDPR). There is no undo and no server-side backup.
• Network exposure. The MCP server is meant for a local client. The optional Streamable HTTP transport ships without authentication — keep it on loopback or behind your own authenticated gateway, or use the hosted service, which puts OAuth 2.1 in front of it.

📚 Full Documentation

All further details, including MCP integration (Claude Desktop), Obsidian sync, Architecture, and the complete API Reference, are available in the official documentation.

👉 Read the Full Documentation