MentisDB 0.10.6.51 — Permanent Skill Delete, Sidecar WAL, Search Persist Wins

August 15, 2026

This release lets operators remove a skill for good — not just revoke it — from the crate, REST, MCP, and the web dashboard. The Skills page now shows how many skills you have and how many are active, revoked, or deprecated. Under the hood, vector sidecars stop rewriting their full JSON snapshot on every append, and MCP/REST no longer hold the chain write lock across that derived I/O.

Upgrade: cargo install mentisdb --locked --force (or let the daemon self-updater install the tag). Restart the daemon after install. Existing sidecars keep working: the next append starts a sibling *.json.wal. Older binaries that only read JSON will treat a sidecar with a pending WAL as stale and rebuild it.

Skills: revoke vs delete

Revoke still keeps every version for audit. Agents should refuse to execute a revoked skill. Permanent delete is the explicit exception to the registry’s append-only lifecycle: SkillRegistry::delete_skill removes the entry, indexes, and latest summary from mentisdb-skills.bin. After delete, list/search/read miss the id and the same skill_id can be uploaded again as Active.

REST: POST /v1/skills/delete with { "skill_id": "…" }
MCP: mentisdb_delete_skill
Dashboard: DELETE /dashboard/api/skills/{id}

Crate, REST, and MCP will delete any existing skill. The dashboard UI matches bearer-token delete: Revoke on active or deprecated rows, then Delete on the revoked list row or Delete Permanently… on the detail pane (type the skill id to confirm).

The Skills list now has a summary bar: total, active, revoked, and deprecated. Those counts come from SkillRegistry::counts() — deleted skills are gone and are not counted.

Vector sidecar WAL

Managed sidecars used to serialize the entire JSON snapshot on every append. Large chains paid O(n) write I/O for one new embedding. 0.10.6 writes one integrity-chained WAL record instead (magic MDBVWAL1, digest SHA-256(prev || entry)). Load verifies the snapshot, then replays *.json.wal. After 32 records the snapshot is compacted and the WAL is deleted.

The append-only thought log is still the source of truth. A corrupt WAL fails verification and the sidecar is rebuilt from the chain.

Search and dashboard persist

MCP/REST append releases the chain write lock before sidecar WAL and overlay I/O so ranked search can proceed. Crate append_thought still flushes derived files before returning.
Implicit-edge first build stays pairwise for N ≤ 128 (exact tests) and uses a temporary HNSW top-k search above that.
Dashboard get_or_open_chain reopens a cached handle only when the on-disk thought count is ahead of the live handle.
query_ranked scores the cached Exact/HNSW backend with search_filtered instead of cloning vectors into a throwaway index.
Exact cosine top-k uses select_nth_unstable_by; adjacency, stemmer, thesaurus slices, HNSW ef_search, webhook Client, and bearer-token registry loads are cached.

Verification

make clippy (warnings as errors)
cargo test --all-features
cargo package --features local-embeddings
Criterion search_ranked smoke (sample-size 10)
Full LoCoMo / LongMemEval not re-run: BM25/vector fusion weights are unchanged. Overlay construction is approximate only for first builds with N > 128.

MentisDB Blog