Garry Tan's AI agent forgot everything between sessions. He built GBrain to fix it. Three use cases for anyone who thinks, reads, and tries to remember things.
I arrived at the same conclusion a few months ago and built my own memory palace with different infrastructure.
It lives in Supabase (Postgres + pgvector) and stores conversation chunks with Voyage 4 Large embeddings (1,024 dims). Search uses HNSW approximate nearest neighbor lookup, followed by Voyage rerank-2.5 cross-attention reranking for precision, across 17 sources
I arrived at the same conclusion a few months ago and built my own memory palace with different infrastructure.
It lives in Supabase (Postgres + pgvector) and stores conversation chunks with Voyage 4 Large embeddings (1,024 dims). Search uses HNSW approximate nearest neighbor lookup, followed by Voyage rerank-2.5 cross-attention reranking for precision, across 17 sources
https://promptedbyeric.substack.com/p/the-blank-page-is-not-deep-work?r=2u5zyx&utm_medium=ios
Nice!