GBrain
GBrain is the durable knowledge layer in a three-layer stack alongside [[projects/gstack|GStack]] (workflow and execution) and [[projects/openclaw|OpenClaw]] (agent runtime). For a comprehensive description of the system, see [[projects/gbrain|the main GBrain page]].
By May 2026, Garry Tan's personal GBrain installation had grown to over 17,000 documents, 4,300 contacts, and 700 companies[^c3], and within one month of its public release expanded to 146,646 pages, 24,585 people, and 5,339 companies — a volume at which Tan stated the system had "become a necessity because grep can't cut it"[^c18][^c19]. The project had accumulated over 16,000 GitHub stars by mid-May, reaching 18,300 stars and 2,500 forks by late May[^c7][^c16]. Tan described GBrain as "the production system I use every day"[^c17].
In June 2026, GBrain released versions v0.42.40.0 through v0.42.52.0 in rapid succession, adding push-based retrieval where the brain volunteers relevant pages without being prompted, brain-resident skillpacks that travel with the repository, git hardening for durable repositories, contention-free sync scaling, and a consolidated autopilot maintenance cycle[^c25][^c26]. In July 2026, version v0.42.56.0 (Life Chronicle) shipped as the project's largest feature expansion to date, introducing a temporal timeline system with event and diary page types and a bi-temporal per-entity ontology[^c27]. The detailed release history is documented in [[projects/gbrain-changelog|the changelog page]].
GBrain follows a [[projects/gbrain-architecture|'thin harness, fat skills' architecture]] using Markdown Git repositories as the source of truth, Postgres with pgvector for hybrid retrieval, and a knowledge graph that auto-wires entity relationships without LLM calls. The system achieved 49.1 percent precision and 97.9 percent recall on its internal [[projects/brainbench|BrainBench]] benchmark. The companion [[projects/gstack|GStack]] project, a collection of role-based slash commands for AI coding, reached 105,000 GitHub stars by June 2026[^c23][^c20]. Tan has described GStack as the "byproduct" of building his earlier work — a working methodology more important than the product that produced it[^c20].
Tan advocates for "tokenmaxxing" — the deliberate willingness to spend aggressively on LLM tokens as costs plummet, treating token expenditure as a competitive advantage — and "skillifying" workflows by packaging every completed task into a reusable skill pack[^c21]. At a hackathon, an agent using this approach judged 85 teams by analyzing code quality, researching participants, watching demo videos, and ranking all submissions in approximately 30 minutes[^c22].
In June 2026, Y Combinator issued a Company Brain Request for Startups identifying domain knowledge as the primary bottleneck for enterprise AI automation, and GBrain was positioned as the key open-source implementation of this concept — a system that transforms organizational knowledge into executable, auditable skill files that AI agents can act upon[^c24]. The open standards ecosystem around the project grew with the formation of the [[concepts/agentic-ai-foundation|Agentic AI Foundation (AAIF)]] under the Linux Foundation, to which [[concepts/mcp|MCP]] was contributed as a founding project. Beyond Tan's own tools, the broader ecosystem includes community extensions bridging GBrain with Notion and other platforms that use GBrain as a downstream knowledge repository[^c15], as well as community ports such as [[projects/rs-gbrain|rs_gbrain]] that reimplement the core architecture in Rust with a SQLite-native storage layer.
GBrain has generated both enthusiasm for its approach to compounding AI memory and debate over its architectural decisions, particularly regarding features that rely on LLM-interpreted instructions rather than deterministic code[^c5]. Later versions have addressed production reliability through self-healing diagnostics, autopilot operation, contention-free database operations, and a sync stall watchdog.