Loomem

What is Loomem?

Loomem is an open-source context layer for LLM agents — a single Rust binary served over MCP that feeds Claude, ChatGPT, Codex, or any MCP client the facts, decisions, and history they need. It is local-first, has no external database, and uses RocksDB + Tantivy embedded.

How to use Loomem?

Install the binary with the install script, then run loomem-server from the ~/.loomem directory. Connect any MCP client, for example Claude, with claude mcp add --transport http loomem http://localhost:3030/mcp.

Key features of Loomem

• One binary with no external services or databases.
• 14 memory_* MCP tools over streamable HTTP.
• Local-first and offline with on-device ONNX embeddings.
• Hybrid retrieval: BM25 + vector + entity-graph weighted fusion.
• Bitemporal facts: distinguishes when a fact was learned vs. when it happened.
• Background consolidation (“dreaming”) merges facts, resolves contradictions, and decays stale entries.

Use cases of Loomem

• Provide LLM agents with a persistent, queryable memory of facts and decisions.
• Enable offline context retrieval without internet access.
• Manage contradictory or temporal facts across multiple conversations.
• Serve as a drop-in memory backend for any MCP-compatible assistant.

FAQ from Loomem

What runtime or database does Loomem require?

Loomem is a single Rust binary with RocksDB and Tantivy embedded. No external database or service is needed.

Is Loomem usable offline?

Yes. The first entry does not need internet because ONNX embeddings run entirely on-device.

What MCP transport does Loomem use?

Loomem exposes its 14 memory_* tools over streamable HTTP at http://localhost:3030/mcp.

How does Loomem handle conflicting facts?

Loomem performs background consolidation (“dreaming”) that merges facts, resolves contradictions, and lets stale entries decay over time.

Where is Loomem’s data stored?

Data resides locally in the ~/.loomem directory created during setup.