GBrain — persistent memory for AI Agents

gbrain search uses tsvector keyword retrieval; gbrain query combines vector similarity, keyword hits and reciprocal rank fusion, returning scored and sourced results such as concepts/do-things-that-dont-scale (score: 0.94) [Source: paulgraham.com]. README benchmark scores are substantially higher than vector-only or keyword-only approaches.

GBrain includes ingestion workflows for meeting transcripts, Email, tweets, voice calls and articles. Import automatically extracts entities and relations, with scheduled overnight entity enrichment. Embeddings offer 14 recipes, including default OpenAI, Voyage, Google Gemini, Azure, local Ollama and llama.cpp, and LiteLLM proxy, supporting cloud or fully local operation.

Use Minions for background work to avoid sub-agent timeouts: gbrain jobs submit <name> triggers deterministic jobs, and gbrain jobs supervisor runs an automatically restarting worker with crash recovery. The README compares roughly 753ms minion completion with sub-agent gateway timeouts often 10 seconds or longer, reporting $0 token cost for this deterministic work and 100% task success.

skills/ contains modular instructions, listed in the README as 30+ skills, for ingestion, enrichment, operations and research; skills/RESOLVER.md dispatches each request to a skill. gbrain skillify scaffold <name> generates 5 stubs for a new skill, gbrain skillpack install <name> deploys curated skill packs, and gbrain check-resolvable audits routing completeness.

GBrain exposes 30+ tools through Model Context Protocol, letting any MCP-compatible client read and write the memory system. For stdio, add { "command": "gbrain", "args": ["serve"] } to ~/.claude/server.json; use gbrain serve --http for production with OAuth 2.1 and a built-in admin dashboard. Subagent execution transcripts are also persisted.

Run git clone https://github.com/garrytan/gbrain.git && cd gbrain && bun install && bun link, then gbrain init to create a local PGLite brain in roughly two seconds. gbrain import ~/notes/ indexes markdown idempotently for safe reruns, gbrain query "your question" performs hybrid retrieval, and gbrain get|put|delete <slug> reads or writes individual pages.

Locally, add "gbrain": { "command": "gbrain", "args": ["serve"] } under mcpServers in ~/.claude/server.json so Claude Code, Cursor, ChatGPT and other MCP-compatible clients can call 30+ tools. Production uses gbrain serve --http with an OAuth 2.1 server and admin dashboard.

The default embedded PGLite requires no external database server. Scale with gbrain migrate --to supabase to Postgres / Supabase, and return with gbrain migrate --to pglite. Bidirectional migration lets you finish locally on a laptop before deciding on cloud deployment, without being locked to either engine.

The 14 embedding recipes include default OpenAI, Voyage, Google Gemini, Azure, local Ollama and llama.cpp, plus LiteLLM proxy for multiple models. Switch providers by choosing a recipe without code changes; fully local operation is available for teams whose data cannot leave their jurisdiction.

gbrain doctor performs health checks with optional automatic repair; gbrain doctor --json suits scripts or CI. gbrain check-resolvable audits routing completeness to confirm every request has a corresponding skill, especially as skillify adds more skills.

In gbrain.yml, set storage.db_tracked for paths such as people/ and companies/ stored in both DB and git, and storage.db_only for media/x/ and meetings/transcripts/ stored only in DB. This keeps transcripts and media from making git excessively large while retaining human-readable core pages.

Submit deterministic, repeatable background work with gbrain jobs submit <name>, then run gbrain jobs supervisor as a worker that restarts automatically and recovers from crashes. Compared with timeout-prone sub-agents, this is faster, has 100% success and incurs $0 token cost for deterministic work.

For existing pages with incomplete graphs, use gbrain extract links --source db and gbrain extract timeline --source db to backfill typed relations and timelines; run gbrain extract all for everything. This relation-building layer uses no LLM, making backfills deterministic and repeatable.

gbrain eval replay checks earlier captured benchmarks for retrieval-quality regressions after code changes. BrainBench-Real stores real queries and scrubs PII for a test set. Integrate it into CI to avoid unnoticed recall regressions during retrieval refactoring.

gbrain import handles initial or full indexing; gbrain sync performs incremental git-to-brain updates. When you or another tool edits markdown directly, sync updates the index and graph without a full rerun. The README-recommended 15-minute sync schedule keeps the brain aligned with git.

llms.txt and llms-full.txt help LLMs establish an overview, AGENTS.md provides the integration protocol for non-Claude Agents, and skills/RESOLVER.md dispatches skills. Start with these three files for the fastest integration into Cursor, Windsurf or a custom Agent.

The lowest-barrier path starts with gbrain init for a local PGLite brain, gbrain import ~/notes/ to index markdown idempotently for safe reruns, and gbrain extract all for typed links and timelines. Then use gbrain query "..." for hybrid retrieval across all notes, and attach gbrain sync for continued synchronization.

gbrain graph-query <slug> traverses typed relations: follow a person's works_at edge to a company and its other people, or a meeting's attended edges to participants. Each page has compiled facts and an append-only timeline; backlink-weighted entity ranking makes frequently referenced nodes easier to retrieve.

Skillify turns manual fixes into reusable capabilities: gbrain skillify scaffold <name> generates a new skill with 5 stubs. Complete content and tests against the conformance checklist, verify routing with gbrain check-resolvable, and optionally deploy curated skill packs through gbrain skillpack install. The same situation then avoids another manual fix.

For teams or multiple Agents, gbrain serve --http exposes memory with OAuth 2.1 and an admin dashboard. Any MCP-compatible client, including Claude Code, Cursor and ChatGPT, can call its 30+ tools to read and write shared knowledge; subagent execution transcripts persist for later review.

Run gbrain query "..." for ranked, sourced hybrid retrieval results, such as concepts/do-things-that-dont-scale (score: 0.94) — PG 關於從不可規模化的努力中學習的論點 [Source: paulgraham.com, 2013-07-01]. Each source makes answers traceable instead of generated without evidence.