Rank #06 · Daily · Agent-Skill
GBrain: Portable Agent Memory You Control Across Harnesses
Give the agent you already use a memory you control — explicit facts with sources, corrections, and the same brain across Claude Code, Codex, Cursor, Grok Bot, and more.
What it does
GBrain is a self-hostable agent brain / memory layer (MIT). GitHub description: "Garry's Opinionated OpenClaw/Hermes Agent Brain." Optional paid hosting at gbrain.io; local install is free.
From the README (as verified today):
- Start keyless with keyword retrieval; optionally add embeddings, reranking, synthesis, and background enrichment.
- Stores explicit facts with provenance; supports corrections and withdrawal; durable preferences can be shared across agents while harness-local state stays local.
- Query modes:
gbrain search(hybrid retrieval of pages) vsgbrain think(synthesized, cited answer plus an explicit note on what the brain does not know yet). - Typed knowledge graph: trusted local page writes can auto-link without LLM calls; breadth-first walks for relationship questions. README cites BrainBench relationship R@5 lift figures — project benchmarks, not TidyHow-reproduced.
- Engines: default PGLite (zero-config) or Postgres + pgvector for larger/shared brains; Markdown repos as the file-backed system of record.
- Install via Bun only (
bun install -g github:garrytan/gbrain#stable) — not the unrelated npmgbrainpackage. Setup helpers for Claude Code, Codex, Cursor, Grok Bot/Muse, OpenClaw, Hermes, ChatGPT, and other MCP clients;gbrain mcp exposefor Tailscale / funnel remote access. - Schema packs (
gbrain-base-v2and others), connectors (Gmail/Calendar, meetings, Slack, voice recipes), cron/"dream" enrichment, and optional Memorable procedural memory (off by default; requires consent).
Author framing: production brain behind OpenClaw/Hermes deployments (author-stated page/person/company counts). Company-brain sharing has explicit grant boundaries — README warns remote grants are not a universal no-leak guarantee.
Who it's for
- Developers who want portable, cited memory across multiple coding/personal agents without locking into one vendor's memory store.
- Operators self-hosting a personal or small-team knowledge brain with MCP.
- People already on OpenClaw / Hermes / Claude Code / Codex who need a durable remember/recall loop.
Less ideal if you only want a chat product's built-in memory with zero install, or if you refuse Bun / self-host ops. Hosted gbrain.io is the lower-ops path.
Stars / language / license
32,157 stars · TypeScript · MIT
Data verified: 2026-11-10 09:10 Asia/Shanghai (GitHub API get_repository + README)
Trending note: +216 stars today (captured 09:06 Asia/Shanghai)
Category tags
Agent-Skill · agent-memory · MCP · knowledge-graph · self-hosted · TypeScript · RAG · OpenClaw
How it differs from similar projects
Compared with vendor-native agent memory (Claude / ChatGPT / Muse built-ins). Those stay inside one product. GBrain's pitch is an explicit, portable record with provenance you can wire to many harnesses via MCP.
Compared with claude-mem / session-compression plugins. Those often optimize compressed context inside one agent session. GBrain emphasizes a durable brain (pages, facts, graph, synthesis) shared across agents and machines.
Compared with plain vector RAG over notes. GBrain adds keyword+semantic hybrid search, schema packs, typed graph walks, cited synthesis with gap analysis, and agent-setup glue — more than an embedding index alone.
Caveats. Not on npm (name collision). Enrichment/synthesis need configured model APIs and cost. Author scale numbers and BrainBench scores are project claims. Company sharing requires careful reading of grant/isolation docs. No TidyHow bake-off against other memory plugins.