CheeseDB is an embedded Go database designed to give AI agents memory with traceable origins and reliable deletion. Every memory must identify the source records or earlier memories it depends on, forming a dependency graph. If information later proves false, malicious, or private, recalling its source removes memories that depend on it, even through multiple steps. In “Fair” mode, memories with an independent supporting source survive; “Strict” mode removes everything the source influenced. Optional embedding models enable meaning-based search and question answering, while forgetting can be requested in plain language and confirmed before deletion.
The engine is append-only and hash-chained to detect silent edits, isolates each agent behind its own token, and offers an erase mode that rewrites files to remove recalled text before returning. Its benchmark, using one million memories per agent, reports that removing 10,000 dependent memories takes 5.2 ms, compared with 68 ms for Neo4j, 236 ms for SQLite, and 1.94 seconds for PostgreSQL; durable write times are broadly similar. CheeseDB trades general-purpose querying for fast provenance-based recall, and keeps its dependency data and search vectors in memory. It is explicitly a preview, has not had an independent security review, and its API and disk format may change. Long startup times for large histories and lack of production guarantees are additional reasons to consider other databases.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.