garrytan/gbrain

★ 15,465TypeScriptAudience · developerComplexity · 3/5Setup · moderate

Mindmap

mindmap
  root((gbrain))
    What it does
      Persistent memory
      Knowledge graph
      Entity linking
      Hybrid retrieval
    Inputs
      Meetings
      Emails
      Notes
      Voice calls
    Retrieval
      Vector search
      Graph traversal
    Integrations
      OpenAI
      Voyage
      Gemini
      Ollama
      llama.cpp
    Audience
      AI agent builders
      Developers

mindmap root((gbrain)) What it does Persistent memory Knowledge graph Entity linking Hybrid retrieval Inputs Meetings Emails Notes Voice calls Retrieval Vector search Graph traversal Integrations OpenAI Voyage Gemini Ollama llama.cpp Audience AI agent builders Developers

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Give a personal AI assistant durable memory so it can recall contacts, meeting notes, and past decisions across separate sessions.

USE CASE 2

Build a knowledge graph of your professional network by feeding GBrain emails and meeting transcripts, then query it by relationship.

USE CASE 3

Integrate GBrain with an autonomous agent platform to add structured long-term memory without extra language-model calls at ingest.

Tech stack

TypeScriptBunPGLiteOpenAIVoyageGoogle GeminiOllama

Getting it running

Difficulty · moderate Time to first run · 30min

Requires a 30-minute setup via agent or CLI plus an embedding provider API key such as OpenAI or Voyage.

In plain English

GBrain is a long-term memory system for AI agents, always-on assistants that read emails, take notes from meetings, or follow your calendar. Out of the box those agents tend to be smart but forgetful: they answer the question in front of them and then lose everything once the conversation ends. GBrain gives them a "brain" that persists, so each new interaction can draw on what the agent has seen before. The brain ingests raw material, meetings, emails, tweets, voice calls, notes, into a local database. As each page is saved, GBrain extracts the people, companies, and other entities mentioned and creates typed links between them (such as "attended," "works_at," "invested_in," "founded," "advises") to build a knowledge graph, without extra calls to a language model. Retrieval uses a hybrid of vector search and that graph, so questions like "who works at Acme AI?" can be answered by following links, not similarity alone. Overnight jobs enrich entities, fix citations, and consolidate memory. The project ships 34 reusable skills, an evaluation harness called BrainBench, and an integration with the LongMemEval benchmark. You would set GBrain up if you run an autonomous AI agent, built on platforms like OpenClaw or Hermes, and want it to accumulate durable, structured memory about your work and contacts. Install takes about thirty minutes via an agent or a standalone command-line tool, the database is PGLite, which needs no server. It is written in TypeScript and runs under Bun, with built-in support for embedding providers including OpenAI, Voyage, Google Gemini, Azure OpenAI, Ollama, and llama.cpp. The full README is longer than what was provided.

Copy-paste prompts

Prompt 1

I want my AI agent to remember past emails and meeting notes. Show me how to set up GBrain with the standalone CLI and ingest a folder of meeting transcripts.

Prompt 2

Using GBrain's knowledge graph, write a query to find all people who work at a specific company based on ingested emails.

Prompt 3

Set up GBrain with Ollama as the embedding provider instead of OpenAI and ingest a batch of plain-text notes from a local folder.

Prompt 4

Explain how GBrain's overnight enrichment jobs work and what they do to entity records between ingest sessions.

Open on GitHub → Explain another repo

← garrytan on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.