mrkeyoor.com_
Tue 11 Aug 13:31 UTC
Self-Hostedevaluationupdated 11 Aug 2026

gbrain

GBrain is a self-hosted memory and knowledge system for AI agents, built to store Markdown, connect related people and topics, retrieve useful context, and turn several sources into a cited answer. It can run locally behind Claude Code or Codex through MCP, or use Postgres, HTTP, OAuth, and per-source access rules as a shared company brain.

Verdict

GBrain is the most complete open-source agent-memory system to consider when plain vector search has become the bottleneck. Its synthesis, graph, Markdown record, MCP surfaces, source scoping, operations tooling, and unusually serious evaluation work justify a pilot. It is not a casual add-on: pin a release, test retrieval on your own corpus, cap provider spending, and choose Postgres before multiple people or agents depend on it.

Setup3/5Fast local start, substantial multi-user and provider operations
Docs4/5Exceptionally detailed, but rapid changes leave some guidance inconsistent
Community5/5Large interest with daily triage, fixes, pull requests, and releases
Maturity3/5Production-minded design, but v0 churn and retrieval defects remain

Who it’s for

Heavy AI-agent users who want durable memory across sessions, tools, people, companies, meetings, and projects.
Claude Code, Codex, Cursor, OpenClaw, or Hermes users who want a local or remote MCP memory service.
Teams prepared to operate Postgres, embeddings, OAuth clients, source scopes, synchronization, and scheduled enrichment.
Knowledge-heavy organizations that value Markdown as the system of record and want retrieval, graph links, citations, and gap analysis in one stack.
Developers who want measurable memory behavior through retrieval evaluations and BrainBench.

Who it’s NOT for

People who only need quick search over a small notes folder: GBrain brings Bun, a database, embeddings, schema packs, synchronization, provider configuration, and more than a hundred MCP operations.
Users expecting the unrelated npm package to work: the README warns that npm install -g gbrain installs something else and may shadow the real binary.
Anyone who needs two local agent clients plus CLI access on the default PGLite brain: issue 3978 confirms that one gbrain serve process owns the database and other processes fail until it exits.
Japanese, Chinese, or Korean teams relying on keyword fallback when embeddings are unavailable: issue 3986 reports that stock Postgres text search leaves the keyword side effectively inert for space-less CJK prose.
Image-heavy teams enabling paid OCR without external spend controls: issue 3973 says routine import and retry paths can call a paid provider per image without a durable aggregate cap.
Applications that require search results to be invariant when only the result limit changes: a current v0.44 report shows the candidate pool changing with the requested limit, causing a page ranked first at five results to disappear at three.

Setup reality

A local trial is easier than the feature list suggests: install from the GitHub repository with Bun, initialize PGLite, run the health check, import Markdown, and connect one MCP client over standard input. Expect an embedding choice and API key unless you deliberately defer embeddings. PGLite's single-process lock makes that path unsuitable for simultaneous clients, and the install guide starts steering larger repositories toward Supabase. The full company setup is infrastructure work: Postgres with pgvector, source design, HTTP MCP, TLS or a tunnel, client registration, least-privilege scopes, database network isolation, synchronization, scheduled jobs, provider budgets, backups, and per-user onboarding.

More than a vector database with an agent label

GBrain is an agent-facing knowledge system, not merely a place to store embeddings. Markdown files in a Git repository remain the record, while PGLite or Postgres supplies search, graph edges, timelines, jobs, access controls, and operational state. Agents can capture material, retrieve raw pages, or ask for a synthesized answer with citations and a note about stale, missing, contradictory, or unsupported information.

That final step is the main reason to consider it. Most retrieval tools return passages and leave the calling model to assemble them. gbrain think combines retrieval and synthesis behind one operation. Automatic links turn Markdown references into typed relationships such as employment, attendance, or investment, allowing multi-hop questions that vector similarity handles poorly.

GBrain also includes ingestion, entity and fact extraction, a durable job queue, background enrichment, contradiction checks, health checks, migrations, and a large agent skillpack. Treat it as knowledge infrastructure, not a weekend plugin.

The local path is genuinely approachable

Installation comes from GitHub through Bun, not the unrelated npm package named gbrain. A trial initializes an embedded PGLite database, imports Markdown, and exposes an MCP server over standard input. Claude Code and Codex connect with one command, and the reduced surface presents five memory verbs instead of the full catalog of 110 operations.

PGLite removes Docker and database administration, but it is a one-process system. One running MCP server owns the data directory. A second Claude session, Codex process, or separate CLI health check cannot open the same brain concurrently. Issue 3978 says the error is clear after the collision, while the client setup guides do not warn users beforehand. Stop the server for maintenance or move to Postgres before connecting several processes.

Embeddings are another dependency. Initialization can detect provider keys, ask the user to choose, or defer setup. Local recipes exist, but the quality path still involves choosing, pricing, and monitoring embedding and reranking services. Provider or dimension changes can also require reindexing.

Shared brains demand real operations

The company design uses Postgres, several sources, HTTP MCP, OAuth clients, scopes, and optional slug-prefix restrictions. The author reports fuzz testing across read paths with zero source leaks. The security documentation states the crucial boundary: source scoping protects HTTP MCP, not a raw database connection. Agent containers must not receive the Postgres password or share a network where they can bypass the application.

The company tutorial estimates about 90 additional minutes after the personal setup, then adds source design, client registration, TLS or tunneling, cron jobs, skills, monitoring, and onboarding. It is useful documentation, but this is not a turnkey multi-tenant service. Operators own backups, upgrades, secrets, database capacity, budgets, permissions, and incident response.

The HTTP server has thoughtful defaults: loopback binding, default-deny CORS, rate limits, hashed tokens, audit logs, and dynamic client registration disabled unless enabled. Internet-facing deployments still need careful reverse-proxy configuration and a protected Postgres network.

Retrieval ambition meets retrieval edge cases

Hybrid search combines vector and keyword retrieval, rank fusion, source boosts, graph signals, reranking, title aliases, and result explanations. The project publishes methodology and evaluation commands rather than relying only on a demo. Release 0.44.0.0 added BrainBench, a hermetic memory test suite with a synthetic corpus, isolation checks, continuity tests, and a CI baseline. Its notes distinguish the production OpenClaw seam from Claude Code and Codex contract-level rows.

Current issues show why corpus-specific evaluation matters. Issue 3982 demonstrates that changing the requested result count changes the inner candidate pool, so a page can rank first at a limit of five and vanish at three. A longer-running ranking issue has several independent reproductions. Issue 3986 explains that stock Postgres tokenization does not give space-less CJK prose a useful keyword fallback, leaving those brains dependent on vectors even when search is described as hybrid.

Build a gold query set from your own notes, languages, page sizes, and naming conventions. Run it on every upgrade, inspect search explanations, and do not equate a green generic benchmark with correct retrieval for your organization.

Automation can create cost as well as value

The overnight dream cycle and ingestion pipelines can keep a brain fresh without manual filing. They can also make provider calls while nobody is watching. Issue 3973 reports that enabling image OCR lets routine imports call a paid OCR provider once per image without a durable total call or spending ceiling. Until a bounded path exists, image-heavy operators should disable that feature or enforce external limits.

Some evaluation and remediation commands do include budgets, health diagnostics, and auditable jobs. Still, each integration needs its own cost review. Retries, large imports, multiple workers, OCR, synthesis, and enrichment can all change the bill.

Extremely active, still moving quickly

GBrain was pushed on August 11, 2026, and release 0.44.0.0 arrived one day earlier. Eleven other releases shown by GitHub landed between August 1 and August 10. The repository had 438 open issues and pull requests, with fixes and detailed reports moving throughout August 11. This is exceptional activity, but the version remains below 1.0 and pinning is essential.

Use GBrain when persistent agent memory is important enough to own as a system. Start with one local brain and five memory verbs. Prove capture, retrieval, citations, costs, and recovery on real data. If those pass, move to Postgres and scoped clients deliberately. Only teams willing to operate the whole brain should adopt the whole brain.

Alternatives

ProjectWhat it isPick it when
Mem0A narrower agent-memory layer with hosted and open-source paths and simpler application APIs.pick this instead when you mainly need to save and retrieve agent memories, not operate a Markdown knowledge system with graph, jobs, and skills.
GraphitiA framework for building temporal knowledge graphs that update as new events arrive.pick this instead when time-aware graph memory is the central problem and you want to assemble your own retrieval and agent surfaces.
KhojA self-hostable personal AI with document answers, agents, scheduled work, and user-facing applications.pick this instead when you want a ready personal assistant experience rather than an agent-operated brain service and schema system.

What people are saying

  1. [github-trending] garrytan/gbrain

Sources

  1. GBrain repository and README
  2. GBrain installation guide
  3. GBrain security guide
  4. GBrain v0.44.0.0 release
  5. PGLite multi-client documentation gap
  6. CJK keyword retrieval limitation
  7. Unbounded paid OCR call report
  8. Hybrid search result-limit inconsistency