gbrain: Embedding permanent, source-attributed memory into AI agents

A tool that grants user-controlled memory capabilities to existing AI agents. It stores explicit facts alongside their sources and supports information modification and retraction. It enables shared memory across multiple agents to prevent context loss.

Unlike search that returns a simple list of pages, it synthesizes stored knowledge to compose actual answers. Each claim is accompanied by its source page, and it provides warnings for information the brain does not yet know or data that has not been updated. This automates complex tasks such as meeting preparation and relationship mapping.

In addition to keyword and semantic search, it supports a typed knowledge graph. It extracts references without LLM calls when writing local pages and enables millisecond-level relationship exploration within dense clusters. On the BrainBench benchmark, the R@5 score for relationship search increased from 0.6626 to 0.7241.

It can be self-hosted under the MIT license, and a paid hosting service is available via gbrain.io. This system, used by Y Combinator CEO Garry Tan in his OpenClaw and Hermes deployments, processes data from over 150,000 pages. It operates as a daemon that collects and refines data around the clock and can be installed within 30 minutes.