A “brain layer” that Garry Tan, CEO of Y Combinator, built to run his own AI agents and open-sourced in April 2026. It is a 24/7 knowledge base that ingests meetings, emails, X posts, voice recordings, PDFs and more, and continuously enriches every person and company it encounters. Its defining trait is that it does not hand back raw search results: it returns synthesized prose with citations, plus a gap analysis of what the brain does not yet know. Knowledge is kept as Markdown files in a Git repository as the system of record, and synced into Postgres for retrieval. It is MIT licensed and can be used from clients such as Claude Code and Codex over MCP.
Key Features
- Synthesized answers with citations: It pulls together people, companies, deals and ideas into prose and makes the supporting sources explicit. It also surfaces the areas where the brain still has no information as gaps, which makes it easier to judge how much to trust an answer
- A self-wiring knowledge graph with no LLM calls: On write, it extracts entity references (people, organizations and so on) from the text and automatically creates typed edges such as
attended,works_atandinvested_in. Because graph construction involves no LLM calls, ingestion cost and latency stay low - Hybrid search: It combines vector search via pgvector, BM25 keyword matching, rank fusion, and boosting based on source tier. The cost/quality balance can be switched across three modes
- A dream cycle that works overnight: Deduplication, citation repair, contradiction detection and preparation of the next day’s tasks all run continuously in the background, so memory keeps getting organized even while you are not using it
- A wide range of ingestion paths: Email and calendar webhooks, voice calls, PDFs and media, ChatGPT / Claude conversation exports, agent session transcripts, Markdown notes, and webhooks from services like Zapier
- Over 100 operations exposed through MCP: Connect from Claude Code, Codex, Cursor, Claude Desktop, OpenClaw / Hermes and others. A lightweight mode exposes only seven verbs such as
recall,rememberandsynthesize, so your tool list stays small - Two database configurations: PGLite (Postgres compiled to WebAssembly) for personal use that runs entirely locally, and Postgres with pgvector for shared team or always-on deployments
Pricing
| Category | Cost | Details |
|---|---|---|
| The software itself | $0 (MIT license) | Every feature is available when self-hosted. With the PGLite setup it can start without any API keys |
| External API usage | Usage-based (per each provider’s rates) | Applies when embeddings, reranking or LLM processing are delegated to hosted services. Not required if you switch to local models |
| Server costs | Actual cost | Always-on and shared team setups are guided at 8GB or more of memory. Personal use is fine running locally |
Pricing is current as of August 2026. Please check the official repository for the latest information.
Pros & Cons
✅ Pros
- Fully open source under the MIT license, with no vendor lock-in
- The system of record is Markdown files, so a human can read the contents and track changes in Git
- Citations and gap analysis make it easier to verify what the agent tells you
- Because graph construction avoids LLM calls, costs stay predictable during large ingestions
- For personal use you can start from a local setup that needs no API keys
⚠️ Cons
- Self-hosting is assumed, so some familiarity with server operations and databases is needed
- Running it always-on means ongoing server costs and API usage fees
- Its design center is a brain for an individual or a small team, so it is not a fit for dropping straight into a multi-tenant product serving many end users
- The personal PGLite setup has a size ceiling; growing past it means migrating to Postgres
- It was released recently, so Japanese-language material and case studies are still scarce
Comparison with Similar Services
| Criteria | GBrain | Mem0 | Zep | Letta |
|---|---|---|---|---|
| Provider | Garry Tan (individual OSS project) | Mem0 | Zep AI | Letta |
| How memory is held | Markdown (Git) plus hybrid search over Postgres | Facts extracted from conversations | A knowledge graph with a temporal dimension | Tiered memory the agent manages itself |
| Primary use case | An always-on brain for a person or team | A memory API embedded in an app | Conversation-first applications | Long-term memory for autonomous agents |
| Delivery | Self-hosted only | Managed plus open source | Managed plus open source | Self-hosted plus cloud |
| Main connection method | MCP (100+ operations) | SDK / API | SDK / API | SDK / API |
Who Is It For
- People who want information scattered across meetings, email and chat gathered in one place for an agent to reference
- People who want memory they can read themselves, kept in a form whose history can be traced in Git
- Developers who use Claude Code or Codex daily and want memory that carries across sessions
- Operators of always-on agents such as OpenClaw or Hermes who are looking for a memory backbone
- Anyone who would rather not hand their knowledge to a SaaS and wants everything to stay on their own server
Summary
GBrain is an open-source foundation for giving AI agents a memory that keeps growing. What stands out is that a whole sequence ─ Markdown as the system of record, a graph and full-text search layered on top, automatic overnight tidying, and answers returned with citations ─ is bundled into a single piece of software. Self-hosting raises the bar for getting started, but for individual developers and small teams who want their memory close at hand, it is a strong option. A good way to try it is to run the PGLite setup locally and ingest the notes and exports you already have.