Most vector databases are built for static corpus retrieval. The indexing is done once and queries are executed forever. Not all systems fit this: agents keep writing their memories, corrections are made to documents, users exercise their right to erasure and sometimes you need to reproduce exactly what a retriever saw two hours ago.
So I created ChronoVec. Internally it uses MVCC just like Postgres, which lets it preserve data for as long as necessary to query "what did this look like before?", but at the same time allows it to actually throw away data when you need to. It doesn't become slower as the data grows, it doesn't consume more memory than necessary, and it can save its state and restore it without losing anything.
When this concept is applied per-branch it results in getting branched operations on top of the vector memory by default: every branch gets its own private delta allowing agents to reason independently and cleanly along different branches.
Most vector databases are built for static corpus retrieval. The indexing is done once and queries are executed forever. Not all systems fit this: agents keep writing their memories, corrections are made to documents, users exercise their right to erasure and sometimes you need to reproduce exactly what a retriever saw two hours ago.
So I created ChronoVec. Internally it uses MVCC just like Postgres, which lets it preserve data for as long as necessary to query "what did this look like before?", but at the same time allows it to actually throw away data when you need to. It doesn't become slower as the data grows, it doesn't consume more memory than necessary, and it can save its state and restore it without losing anything.
When this concept is applied per-branch it results in getting branched operations on top of the vector memory by default: every branch gets its own private delta allowing agents to reason independently and cleanly along different branches.