Manage personal knowledge by letting AI agents organize memory and provide sourced answers
When you ask about a contact for a meeting tomorrow, you can receive specific answers with cited sources and notes on what is still unknown, rather than just a list of pages. GBrain adds a user-controlled memory function to the AI agents you already use. It does not just search; it reads and synthesizes stored information, and it organizes and supplements memory over time.
Providing answers instead of search results
Most personal knowledge tools return a list of pages based on keywords, but GBrain generates sentence-form answers that synthesize stored content via the gbrain think command. These answers specify the source for each claim and note if the data is outdated or contradictory. For example, it summarizes the last conversation with a specific person and unresolved tasks, while pointing out information gaps that the brain has not yet recognized so you can verify them yourself. This process is like a brain reading pages and writing an answer, unlike a search engine finding pages.
A schema that reflects your file structure
GBrain does not force a fixed folder structure; instead, it provides a schema pack system that suggests types based on your existing file system. You only need to give final approval using gbrain schema review-candidates. The default gbrain-base-v2 pack supports 15 standard types, including people, companies, emails, and meeting notes, and you can create custom packs to import existing notes or data from Notion and Obsidian as is.
An automation loop that refines memory while you sleep
GBrain uses cron jobs to maintain memory even when you are not active. This process merges duplicate person pages, fixes citation errors, scores information importance, finds contradictions, and prepares tasks for the next day. When writing local pages, it automatically generates graph links using pattern matching without LLM calls, connecting only information from trusted sources to the graph. This background processing ensures the agent always responds based on up-to-date memory.
Things to check before deciding on installation and operation
GBrain is distributed under the MIT license and is free for self-hosting. Installation requires Bun 1.4.4 or higher, and you must install it directly from the GitHub repository rather than as an npm package. You can start without a key, but using semantic search or synthesized answer features may require separate model API configuration and incur costs. Additionally, if you configure cloud-based embedding or reranking providers, text may be sent to those services, so you must verify data boundaries when handling sensitive information. If you use company sharing features, you must clearly set access permissions and visibility filters for authenticated remote clients, and recognize that local files and shared database credentials belong to different trust boundaries.
