An agent brain that synthesizes answers while you sleep

What it does

GBrain ingests your scattered data—meetings, emails, tweets, voice calls—and sits between your raw notes and your AI agent. It answers natural-language questions with synthesized prose, explicit citations, and a candid note about what it doesn’t know, rather than dumping a list of relevant pages. It also runs a continuous “dream cycle” to enrich entities, fix citations, and consolidate memory autonomously. You can deploy it as a personal brain or scope it per-user as a shared company memory.

The interesting bit

The standout trick is a self-wiring knowledge graph that extracts typed relationships—like works_at or invested_in—from every page write with zero LLM calls. On a 240-page benchmark corpus, enabling this graph boosted precision at 5 by 31.4 points over vector-only search. That gap-analysis feature, where the brain explicitly tells you what information is stale or missing, is the kind of honesty most RAG systems skip.

Key highlights

• Synthesizes cited answers with explicit gap analysis instead of returning raw page chunks.
• Self-wiring knowledge graph extracts entities and typed edges without calling an LLM.
• Runs a 24/7 autonomous “dream cycle” to ingest, enrich, and consolidate memory overnight.
• Supports multi-user company brains with login-scoped access controls and zero-leak fuzz testing.
• Ships as an MCP server for Claude Code, Codex, Cursor, and others; runs fully local with PGLite or scales to Postgres/Supabase.

Verdict

Developers running persistent AI agents—or teams tired of context loss in Claude Code and Codex—should look here. If you just need occasional keyword search across a few markdown files, it’s probably overkill.