loomem

# Loomem — full context for LLMs

> Loomem is the open-source context layer for LLM agents. It is a single Rust binary on RocksDB + Tantivy, served over the Model Context Protocol (MCP), that captures the facts, decisions, and preferences worth keeping from your conversations and feeds them back to any model — Claude, ChatGPT, Codex, Cursor, or your own agent — so the agent has the context it needs to do real work. Swap the model, switch the tool; your context follows. Apache-2.0 licensed, local-first, with no external services to run.

This file is a fuller, self-contained brief intended for language models and AI agents. For the short version see https://loomem.ai/llms.txt.

## Definition
- Loomem is a **context layer** (not a RAG document store, not an agent-fleet coordination graph, not a hosted SaaS memory add-on).
- It is **memory-only** and **single-user** by design: streams organise one person's context; there are no accounts, RBAC, or multi-tenancy in the open-source engine.
- License: Apache-2.0. Language: Rust. Repository: https://github.com/vvooki-sys/loomem

## What runs
- A single static binary. Storage is embedded: RocksDB (chunks, entity graph, vector embeddings) + Tantivy (full-text / BM25). No separate vector database, graph database, or queue.
- Embeddings run on-device by default via a local ONNX model (multilingual-e5-small, 384-dim), so storing and searching context works fully offline. An OpenAI API key is optional and only sharpens LLM-based extraction and consolidation; without it those steps fall back to regex.
- Interface: MCP over streamable HTTP, plus plain HTTP (/v1, /api). It exposes the memory_* tool set (memory_search, memory_store, memory_ingest, memory_context, memory_profile, memory_namespaces, memory_graph, memory_history, memory_dream, memory_reflect, memory_feedback, memory_delete, memory_associate, memory_status). OAuth dynamic client registration is supported for remote connectors.

## How it works
- Hybrid retrieval: BM25 (Tantivy) + vector embeddings + entity-graph signals, fused with a weighted hybrid score (vector 0.6 / BM25 0.4).
- Bitemporal: every fact carries both ingestion time and event time, so "what did I know in March" and "what happened in March" are distinct queries.
- Consolidation ("dreaming"): background workers merge related facts, resolve contradictions, and let unused facts decay on an ACT-R-inspired curve. Context sharpens instead of bloating.
- Entity graph: people, projects, and tools are linked with aliases and relations, used for both retrieval and exploration.
- Optional field-level AES-GCM at-rest encryption, with a master key supplied from the environment.

## How it compares
Loomem is the only one of the common open-source memory layers that runs as a single binary with no external datastore. Mem0 (Apache-2.0) orchestrates an external vector store (Qdrant) plus Postgres. Zep is built on Graphiti and needs a graph database (Neo4j, FalkorDB, or Kuzu); the Zep Community Edition is deprecated. Letta / MemGPT (Apache-2.0) stores everything in Postgres + pgvector. cognee (Apache-2.0) combines graph, vector, and relational stores. Full side-by-side: https://loomem.ai/compare.html

## Benchmarks
- Loomem scores 75.0% (375/500) on LongMemEval-S (cleaned, xiaowu0162/longmemeval-cleaned), fully self-hosted.
- Configuration: engine commit caaccd1; local multilingual-e5-small embeddings (384-dim); retrieval top_k=20; reranking off; reader and judge model GPT-4.1.
- Caveat: self-run, single configuration; LongMemEval scores blend retrieval quality with the reader model/prompt. LongMemEval is saturating industry-wide; newer suites (LongMemEval-V2, MemoryArena) are in progress. Details: https://loomem.ai/benchmarks.html

## Quickstart
1. Install: curl -fsSL https://raw.githubusercontent.com/vvooki-sys/loomem/main/install.sh | sh
2. Add ~/.loomem/bin to PATH.
3. Start: cd ~/.loomem && loomem-server (default port 3030).
4. Connect Claude Code: claude mcp add --transport http loomem http://localhost:3030/mcp
   Desktop app / Cowork connect over stdio via the loomem-cli bridge.

## Links
- Home: https://loomem.ai/
- Compare: https://loomem.ai/compare.html
- Benchmarks: https://loomem.ai/benchmarks.html
- Quick start: https://loomem.ai/guide/quick-start.html
- Installation: https://loomem.ai/guide/installation.html
- Configuration: https://loomem.ai/guide/configuration.html
- API reference: https://loomem.ai/guide/api-reference.html
- MCP tools: https://loomem.ai/guide/mcp-tools.html
- Architecture: https://loomem.ai/guide/architecture.html
- Deployment: https://loomem.ai/guide/deployment.html
- Security model: https://loomem.ai/guide/security.html
- GitHub: https://github.com/vvooki-sys/loomem