An agent working from a partial view of a codebase writes code that drifts from its patterns, ignores decisions made for a reason, and misses the callers a change breaks. That is technical debt someone pays for later. NeuralMind gives the agent the project’s structure and history, and checks changes against them. Here is what does that, and what recent releases added.
Latest release
v4.24.0Released 2026-11-02Each works on your machine, from your repository, and each has a command that shows it.
neuralmind drift maps your staged diff onto the code graph, groups each changed function with its siblings, and flags one that leaves out something most of them do: the tenth handler that forgot the auth check the other nine make. Only code in the diff is judged. neuralmind init-hook wires it into pre-commit; it warns by default, and --strict blocks.
$ neuralmind drift . --staged
Record why the code is the way it is, with the files the decision governs. When a commit changes one of those files the decision is marked STALE, and the agent is warned before it edits. A reversed decision can be retracted so it stops reaching the agent (v4.16.0+).
$ neuralmind decisions record --title "Hash tokens before storing" --rationale "Leaked DB dumps" --files auth/tokens.py
neuralmind impact lists every caller, importer and subclass of a symbol from the static code graph, so a rename or a signature change reaches all of them instead of breaking the ones nobody remembered.
$ neuralmind impact verify_session
neuralmind review reads your diff and lists files that have been edited together with the ones you changed but are not in this diff, from what the synapse layer learned on your project.
$ neuralmind review . --base main
Answers from a stale index carry one line saying so: when commits since the last build changed the project, or the background rebuild failed. Since v4.13.0 the index also rebuilds after pulls, merges, branch switches and rebases, not only after your own commits.
$ neuralmind init-hook .
v4.16 to v4.23 at a glance, then every release, newest first, with its full notes.
Timed on one machine, with the output checked byte for byte. The release notes give the numbers.
Measured offline, not in agent runs. The release notes give the numbers.
security.profile: cuiv4.16.0neuralmind hardening checks a 17-rule baselinev4.19.0neuralmind data-locations lists where the data isv4.18.0These support controls in your system. Your assessor decides compliance.
40 of 40
public-benchmark queries where NeuralMind found the file the fix changed. Keyword search with ripgrep: 32 of 40.
Reproducible0 of 60
pilot runs in which the agent called a NeuralMind tool. Cost per task showed no measurable difference: +0.0% (95% CI −5.2% to +8.6%).
PilotOn a larger eval the right file is in the top five on 140 of 150 questions, and all 10 misses are published. Getting agents to use what NeuralMind finds is the work now.
graph.json a full build writes, byte for byte. Before, a rebuild could link a call to a different function of the same name, and files .gitignore and .neuralmindignore exclude counted as new on every build.neuralmind watch --reindex follows the same rules..neuralmind/type_cache.json, keyed by a hash of its text, so a rebuild parses only the files that changed, with the same type edges.ctx.session_cwd()), so a session in a built checkout gets its context. Older Hermes is unchanged.NEURALMIND_PROJECT counts only in the gateway’s own environment, where it applies to every session like a pin.test_x.py tests x.py.neuralmind doctor opens with a Wiring section (daemon, Git hooks, Hermes plugin, MCP registration per client), and each warning carries the command that fixes it.neuralmind install-mcp --client hermes registers the MCP server with Hermes-Agent, and a build ends with the commands for whatever isn’t wired yet.main.initialize reply tells the agent when to call neuralmind_search before Grep, and tools/list offers six tools by default instead of 28.cryptography. On Windows on ARM, use WSL 2 or the container image.neuralmind hardening, which checks a host against a 17-rule baseline.neuralmind demo runs offline, straight after pip install.neuralmind data-locations lists everything NeuralMind writes for a project, in it and outside it, with the command that removes each.security.profile: cui turns on OS-account identity, encrypted-storage checks and loopback-only serving together, and signed checkpoints make an audit log you can’t silently shorten.neuralmind eval . --synapse-ab runs your own questions with the synapse layer on and off and reports what it moved.Older releases: the full history.
Every number, with its source and the command that reproduces it
Free at one seat, local, and no account. Build the index, wire the hooks, and the guards above run on your next commit.
$ pip install neuralmind
$ neuralmind build .
$ neuralmind init-hook .