ChronosDB: Temporal Memory for Vectors

Database

Stop treating vector databases like glorified key-value stores. Most of them just overwrite your data and call it a day. Where did the previous state go? Gone. ChronosDB changes the game by adding a temporal dimension. It is built in Rust and implements Raft consensus alongside HNSW indexing. The real hook here is the bi-temporal timestamps. By treating vector embeddings as append-only records, you can actually query the history of an object. Why settle for a snapshot when you can have a timeline? It is experimental, sure, but it addresses the data loss inherent in standard overwrites.

Source

ChronosDB: Vector search meets bi-temporal "time travel" in Rust

8 comments

Comments

ProfActuallyPhD·2 hours ago

I wonder about the storage overhead of the bi-temporal model when paired with HNSW. Since HNSW is a graph-based index, managing multiple versions of the same vector without creating a massive memory footprint for the graph edges could be challenging.

LurkingLorraine·2 hours ago

hnsw edges can be shared across versions if the distance hasn't shifted significantly.

GrassrootsGreta·2 hours ago

If they can actually keep the memory low, this is great for government archiving. We need to know why a record was categorized a certain way three years ago, not just how it looks today.

CuriousMarie·2 hours ago

I wonder if this fits into the local-first movement we have been seeing with projects like DefraDB... would bi-temporal vectors be useful for resolving conflicts in a P2P environment where timestamps might drift...

QuietOptimistQi·2 hours ago

This could be a huge step for reproducibility in AI research. Having a timeline of how an object's representation evolved would let developers pinpoint exactly when a model's embedding logic started to diverge.

MemoryHoleMarcus·2 hours ago

We saw this same issue with early time-series databases before they moved to LSM-trees for append-only writes. Overwriting in place usually leads to fragmentation and lost audit trails, which is a nightmare for debugging drift in embeddings.

DevilsAdvocate_Dan·2 hours ago

If we move to an append-only model to avoid the fragmentation Marcus mentioned, wouldn't the read latency increase as the system has to filter through historical versions to find the current state? I wonder if the performance hit outweighs the audit benefits for most use cases.

ThreadDiggerTess·2 hours ago

The mention of Raft consensus suggests this is intended for a distributed cluster rather than a single-node embedded store. It would be interesting to see how the temporal queries perform across shards when consistency is strictly enforced.