GBrain is the knowledge infrastructure Garry Tan describes as an extension of Karpathy's llm-wiki-pattern into a personal brain repository. In Tan's account, it stores roughly 100,000 structured pages covering people, companies, meetings, books, articles, podcasts, videos, and ideas.source: garry-tan-meta-meta-prompting-ai-agents-2026.md
The page schema he describes has compiled truth at the top, an append-only timeline below, and raw data sidecars for source material. After meetings, entity propagation updates the pages for every person and company mentioned, making the system closer to a nervous system than a filing cabinet.source: garry-tan-meta-meta-prompting-ai-agents-2026.md
GBrain is positioned as the fat data layer for personal-agents: the harness can route tasks, and skills can execute workflows, but the compounding advantage comes from accumulated, cross-referenced personal context. Tan says GBrain ships installable skills and reports strong retrieval performance on LongMemEval, but this wiki currently has only his article as the source for those specific claims.source: garry-tan-meta-meta-prompting-ai-agents-2026.md
Related pages: garry-tan, skillification, self-improving-knowledge-base, llm-wiki-pattern, harness-engineering.