NeuralMind v4.5.1 — decision memory answers questions, for every role

Two fixes to decision memory, both meant for v4.5.0, that landed just after it was tagged:

1. Decision search answers questions. A question used to find nothing unless every one of its words appeared in a decision.
2. The default MCP roles reach the memory retrieval tools. neuralmind_memory_search, neuralmind_memory_timeline and neuralmind_memory_get returned security_denied to every caller that didn’t declare admin.

1. Decision search answers questions

Decision search required every word of the query, so neuralmind decisions query found a decision for “sqlite wal” but not for “how do we handle sqlite wal?”, and agents send questions.

• Any word of the query can match. Common words such as “how” and “the” are ignored, and decisions matching more of the words, and rarer ones, rank first. This covers decisions query, the neuralmind_query_decisions and neuralmind_memory_search MCP tools, and the LIKE fallback used when SQLite lacks FTS5. A question nothing answers can still return partial matches, so check the titles.
• Status filters are case-insensitive, and ALL returns every status (over MCP, the advertised "all" used to match nothing). The CLI’s --status also accepts INVALIDATED. An unknown status is an error, invalid_request over MCP, instead of an empty result.
• neuralmind decisions eval leaves your decisions alone. It used to delete the project’s .neuralmind/memory.db and leave synthetic decisions with the author eval-harness behind; it now runs on a scratch store. The Memory Layer wiki shows how to retire any it left. --format md no longer crashes.
• neuralmind decisions eval --queries FILE scores search against questions with known answers: recall@k and MRR as mean and range per query kind, with every miss and false positive listed. The committed query set and its measured results are in the Memory Layer wiki.
• MCP errors say what to do. A missing index is code: "index_not_built" with a hint to call neuralmind_build, not security_denied; security_denied carries reason: "rbac" or "rate_limit". See Troubleshooting.

2. The default roles can use the memory retrieval tools

neuralmind_memory_search, neuralmind_memory_timeline and neuralmind_memory_get were never added to the default role policy. Through the MCP server, a caller that declared no role (which means builder) or declared reader got security_denied from all three, so agents couldn’t use the progressive-retrieval path the Memory Layer wiki describes.

• Builder and reader now have all three. They only read.
• Recording and invalidating decisions stays builder-only.
• synaptic_neighbors, structural_neighbors, next_likely, impact and review stay admin-only by default, as the Security Guide lists.
• A test now requires every MCP tool to be granted to the default roles or listed as left out on purpose, so a new tool can’t ship outside the policy unnoticed.

A security.roles policy in neuralmind-backend.yaml replaces the defaults, so a project that sets one keeps exactly the tools it lists.

What the agent actually sees post-install

Upgrade notes

• No reinstall and no configuration change: pip install -U neuralmind.
• MCP error codes: a missing index returns index_not_built instead of security_denied, and security_denied now carries reason (rbac or rate_limit). A client that matched security_denied to mean “not built” needs updating.
• If you passed role: "admin" only to reach the memory retrieval tools, you can drop it.
• If neuralmind decisions eval ever ran inside one of your projects, look for decisions with the author eval-harness; the Memory Layer wiki shows how to retire them.

Related

• CLI reference: decisions
• Memory Layer wiki · Security Guide: access control
• Previous release: v4.5.0