highascanbe

openclaw stores memory as markdown files and searches them with sqlite vector embeddings and BM25 keyword matching. it works but thats a whole RAG pipeline running locally for every recall nuggets does something different. instead of chunking text and computing embeddings it compresses each fact into a complex valued holographic vector using HRR math. multiple facts superpose into one fixed size object and recall is just an algebraic unbind operation openclaw needs sqlite extensions fts5 and sqlite-vec to search. nuggets needs zero dependencies. openclaw chunks and embeds markdown. nuggets binds key value pairs into phase angles the result is sub millisecond recall with no database no embedding model and no file parsing. facts that get recalled often auto promote to permanent context so the agent gets cheaper over time not more expensive both are local first and open source. different math same goal

Nuggets plan is actually insane

Appreciate the community stepping up to solve the 'Rented Cognition' problem. I don't want VCs; I want recurring usefulness. $NUGGETS is officially the research grant for the Agent’s autonomy. Back to the lab to get this Mac Mini setup. 🧪🍗