GBrain

GBrain is a product that searches a large markdown corpus with hybrid retrieval and a link graph.¹² Keyword and vector search run together, and their results are fused, reranked, and returned as pages.³

Overview#

Wiki links can be extracted into graph edges at write time without a model.³ A sourced answer is a different mode from raw search.³ Published recall figures are the vendor’s own and remain claims until measured independently.⁴

Mechanism#

1. Keyword and vector search run in parallel.³
2. The ranks are fused and reranked, and the top pages are taken.³
3. Wiki links are extracted into a graph when a page is written.³
4. A separate mode answers with citations and names gaps.³
5. Retrieval is held to regression and isolation checks before it is trusted.³

Applications#

The product is relevant when deciding whether hybrid search plus a link graph is worth measuring on a given wiki.

See also#

• Shared brain – the memory layer
• Avid memory stack – measure-first memory
• Karpathy LLM wiki foundation – the compiled wiki
• Block Buzz – an adjacent tool note
• Thin harness, fat skills – the thin harness
• DataClaw – a data-science harness

References#

Footnotes#

1. https://github.com/garrytan/gbrain ↩

2. https://x.com/garrytan/status/2042497872114090069 ↩

3. https://x.com/garrytan/status/2058448209027141709 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

4. https://github.com/garrytan/gbrain-evals/blob/main/docs/benchmarks/2026-05-23-v0.40.6.0-snapshot.md ↩