LabscoMCP early access ↗
Bobby-cell-commitsMCP SERVER

Open-Brain

Self-hosted memory that stores thoughts as a weighted entity graph, with a background cycle that deduplicates, synthesises and archives.

by Bobby-cell-commits ↗Agent MemoryCommunity

LABSCO SUMMARY

It publishes a benchmark number, which most memory servers do not.

37.2% on LongMemEval is a figure you can argue with, and that is the point — the alternative is a memory system that makes no measurable claim at all. The design behind it is the graph: connections above 0.80 similarity get typed by an LLM, so "contradicts" and "supersedes" are stored relations rather than something re-derived at query time, and co-occurrence edges strengthen links that keep being retrieved together. Running it means running Postgres with pgvector and paying OpenRouter for embeddings.

What it is

A memory server you host yourself. Thoughts arrive from Telegram, RSS and paper pipelines, or straight from an AI client, and land in an entity graph rather than a flat document store — embedded, deduplicated, auto-linked and entity-resolved on the way in.

What you get

  • capture_thought saves a new thought with automatic embedding; search_thoughts does semantic search with optional one-hop graph expansion
  • list_thoughts browses by type, topic, person, theme, quality and time; thought_stats returns counts and breakdowns
  • get_connections traverses typed links — extends, contradicts, is-evidence-for, supersedes — classified by an LLM above 0.80 similarity
  • list_entities browses the people, tools, projects and organizations extracted from your notes
  • weekly_review summarises themes, open loops and next steps; analyze reports hubs, density, sources, co-occurrence and themes
  • Maintenance is exposed rather than hidden: dedup_review with a similarity histogram, refresh_salience, review_stale, serendipity_digest for forgotten high-quality notes, and pipeline for ingest health
  • update_thought re-embeds and re-extracts metadata; delete_thought cascades connections; migration_guide covers importing from other platforms

Requirements

Either Supabase or Docker Compose, plus PostgreSQL with pgvector and an OpenRouter API key for embeddings and metadata extraction. Clients connect over streamable HTTP to the deployment's /open-brain-mcp/mcp path with your key in an x-brain-key header — for Claude Code, claude mcp add --transport http --header "x-brain-key: ...". A Telegram bot is optional and is the intended capture path on the go.

Setup effort

Follow the published steps — npm install -g supabase

See installation options