RuVector 0.2.40 exposes its research control plane through npm, CLI and MCP

What RuVector is

Most vector databases return the same answer forever. RuVector stores embeddings, answers graph queries, and adjusts its own index based on what actually turned out to be useful.

What changed in 0.2.40

The release publishes the optimisation control plane — MetaHarness, Darwin and Flywheel — through the npm SDK, the CLI and the MCP server. Dependencies are pinned directly so an installed package has the capability set that was tested.

Work has continued since: in early September the project recorded a rejected research direction, mincut gated forgetting, with the evidence for the rejection written down as an ADR. Negative results getting the same treatment as features is the useful part.

Get started

Node.js 20 or newer.

npm install ruvector@0.2.40

Check what the installed package can actually do:

npx ruvector harness doctor --json

Expected result: a JSON report loading and checking the nine pinned capabilities, including the router, safety controls and workspace lens.

For project scoped agent memory:

npx ruvector hooks remember --semantic --type decision \
  --text "we chose SQLite over Postgres for the local cache"

The first semantic command downloads and caches a local embedding model. Keep one model and dimension per store, and use npx ruvector hooks reembed before changing an existing store.

Use it today

Practical case: a decision log your coding agent can actually search. Input is a one line decision at the moment you make it; the workflow is remember during work and recall at the start of the next session; the output is a ranked set of prior decisions instead of a guess.

npx ruvector hooks recall --semantic --top-k 3 --query "local cache"

Acceptance test: store three decisions, then recall with wording that matches none of them literally. Semantic mode should still return the right one.

Push it further

Experimental commentary. The exotic end is a database that rewrites its own retrieval policy overnight and can prove the new policy is better. The dependency is measurement: self improvement without an honest gate is just drift.

Falsifiable test: freeze the index, record recall quality on a fixed query set, let it learn, then re-measure on the same set. If the number does not move, the loop is not doing anything for your data.