[ruvnet]

Summary

npx ruvector now has create/read/update/delete on every data surface: vector stores, RVF stores and hook memories, through the CLI, the MCP server and the library.

Surface	Create	Read	Update	Delete
Vector DB — CLI	create, insert	get, search, stats	update	delete
Vector DB — MCP	vectordb_create, vectordb_insert	vectordb_get, vectordb_search, vectordb_stats	vectordb_update	vectordb_delete
Vector DB — library	insert, insertBatch (now true upserts)	get, search	update	delete, deleteMany
RVF — CLI / MCP	create, ingest (upsert)	query, status	ingest replaces an id	rvf delete / rvf_delete
Hook memory — CLI	remember	recall (now with ids), memory-get, memory-list	memory-update	forget
Hook memory — MCP	hooks_remember	hooks_recall, hooks_memory_get, hooks_memory_list	hooks_memory_update	hooks_forget

The native VectorDB has no enumerate API, so there is no vector list; get by id is the read.

Bug fixed: duplicate search results after re-inserting an id

Re-inserting an existing id left the old node in the native HNSW index, so search returned that id twice, once with its stale vector. Duplicate ids within one insertBatch did the same. insert and insertBatch are now true upserts (the last entry wins). The root cause is in the native core; this is a wrapper-level fix that avoids a multi-platform native republish.

Memory ids

• MCP memories had no id.
• CLI ids (mem_<ms>) could collide within one millisecond.
• MCP's legacy conversion gave every id-less memory the same id, which collapsed them in the engine.
• New ids are collision-free.
• Legacy entries get deterministic ids, so the CLI and the MCP server derive the same ones.
• The MCP engine's recall set follows edits and deletes, including ones made by another process.

Security

• Shared validation in bin/record-validate.js, used by both the CLI and MCP:
  • ids: bounded length, no control characters
  • vectors: finite numbers of the store's dimension
  • metadata: a plain object with no __proto__/constructor/prototype keys at any depth; ≤64 KB, ≤8 levels
  • batches: ≤10k entries; k ≤ 1000
• vectordb_* paths are confined to the project (RUVECTOR_MCP_ROOTS to widen) and are always explicit, so the ./ruvector.db default can never be created from MCP.
• Memory updates re-embed through the same embedder and ADR-210 provenance gate as remember, so an edit cannot create a mixed store.
• The readonly MCP profile gains only hooks_memory_get and hooks_memory_list. All new write tools stay hidden and can't be called.
• RVF tools:
  • rvf_ingest, rvf_query and rvf_delete now validate their input.
  • rvf_open no longer leaks the store handle.
  • rvf_delete accepts string ids; it previously accepted only numbers, while the store uses string ids.

Tests

tests/crud.test.mjs: 9 end-to-end tests covering the library, CLI subprocesses and MCP over stdio, including rejection cases. 7 of them fail on 0.3.5; the other two are a unit test of the new id module and the readonly-profile check, which have nothing to compare against on 0.3.5. Locally, npm test passes and node --test tests/*.test.mjs passes 61/61.

Benchmark

scripts/bench/crud-bench.mjs (not shipped). Run: 10k × 384-dim, native, Linux x64. Values are p50 / p95.

op	p50	p95
get	11.5 µs	14.7 µs
insert, new id	2.94 ms	4.25 ms
insert, existing id (upsert)	2.98 ms	4.04 ms
update, vector	2.68 ms	3.97 ms
update, metadata merge	2.57 ms	3.69 ms
delete	144 µs	195 µs
search, k=10	1.45 ms	2.23 ms
memory get / list / forget (5k store)	3.5 / 16 / 57 µs	10 / 100 / 126 µs

• The upsert check costs one get (~12 µs) per id, under 0.5% of an insert, and is skipped entirely when loading into an empty store.
• Backfilling ids on a 5k-memory store takes 8.6 ms, once.

🤖 Generated with claude-flow