Your team fixes a logout bug.
They keep sign-ins on the server so a stolen key stops working as soon as someone logs out.
Hunch keeps decisions, completed work, and commitments with their sources. Start with engineering memory for coding agents, or connect agents to a self-hosted state server.
It sees files, not why decisions were made, what failed before, or what a small change could break.
They keep sign-ins on the server so a stolen key stops working as soon as someone logs out.
The old discussion is hard to find. The safer code now looks more complicated than it needs to be.
The change looks clean, but it brings the same logout bug back.
With the lesson recorded and the integration active, Hunch can show the decision and the bug it prevents before the edit.
Hunch keeps structured records in Git. Inspect them in a browser or connect agents through MCP, HTTP, the CLI, TypeScript or Python.
memorySave decisions, rejected approaches, bugs, and open questions alongside the code they explain.
decisionsConflicting current decisions are refused. Replacing a decision leaves a history of what changed.
accessChoose which agents can access a partition, then narrow access to individual records when needed. Optional key-bound credentials add request signing.
actionsInspect completed work, open commitments and cited sources in the browser. Unknown or unverified work keeps that status.
controlRecord sourced user, team and organization conventions. They stay advisory; blocking requires an explicitly trusted rule and strict mode.
gitRecords are readable files. Git lets you review and revert committed changes; indexes can be rebuilt.
Hunch gives the assistant a focused brief and tools to check its work. The assistant still plans and carries out the task.
Capture decisions, corrections, and findings. Link them to code and sources so the reason remains findable.
Select relevant memory for a task through MCP and supported assistant hooks. Coverage depends on the integration.
Evaluate supported rules and keep a task report of delivered memory, reported use, and observed checks.
Task reports show the memory delivered, what the agent says it used, and the results of rules and commands that actually ran.
npm i -g @davesheffer/hunch
A passing test alone does not prove Hunch caused the result. Missing evidence stays unverified. Read a task report →
Engineering memory connects the reasons behind code to its dependencies and the rules the team chose to protect.
whyRead decisions, rejected ideas, and past failures behind a file or symbol.
blast radiusFollow indexed code relationships to understand what a change could affect.
project DNAInspect observed terminology and working habits. These observations stay advisory.
compareCompare candidate branches against recorded decisions and constraints.
conformCheck supported code relationships, such as whether a payment path still calls authorization.
change gateFlag conflicts with supported rules. Strict mode blocks only where the rule and integration permit it.
Want the technical details? Read how Hunch works →
Install the CLI and run hunch init inside the project you want it to remember.
Optionally backfill recent Git history, then review the captured decisions and sources.
Reload your assistant and ask why a file is built this way. In Codex, trust the hooks through /hooks first.
# install from npm — Node 22.13+ $ npm install -g @davesheffer/hunch # connect Hunch to this project $ cd your-repo && hunch init # learn from the last 90 days $ hunch backfill --since 90d # see your project's DNA $ hunch dna inspect # ask what a file is for $ hunch why src/auth/session.ts
npm i -g @davesheffer/hunch && hunch init
Hunch gives advice by default. Blocking requires your explicit choice. To update, run hunch update, reconnect the assistant, and review changed Codex commands in /hooks.
Claude Code? Install as a plugin instead:
/plugin marketplace add davesheffer/hunch → /plugin install hunch@hunch
Open VSX
The goal: a customer issue becomes a code fix and a verified follow-up. Shared state ships today; the Sofia pilot is testing the complete handoff.