Deterministic StateWhat deterministic state means, in plain words.Read the thesis →

Your agents forget.
Hunch remembers.

Every AI session starts from zero. Hunch keeps what your team decided, what got done and what’s still owed — with sources, in Git — so the next agent picks up where the last one left off.

Get started See how it works
the missing context

AI reads your code. It wasn’t there when you decided.

It sees what the code does — not why, what broke last time, or what a small change will break next.

Monday

Your team fixes a logout bug.

Sign-ins move to the server, so a stolen key dies the moment someone logs out.

Months later

The code stays. The reason doesn’t.

The discussion is buried. The safe code now just looks overcomplicated.

Next session

An AI “cleans it up.”

The diff looks great. The logout bug is back.

With Hunch

The reason shows up first.

Once the lesson is recorded, Hunch puts the decision — and the bug it prevents — in front of the agent before it edits the file.

Hunch Memory in 2 minutes. Then go deeper in the 7-episode series: setup, hooks, capture, delivery, rules, drift and teams.
available today

One record. Every session. Every agent.

Plain files in Git. Browse them in your browser, or plug agents in through MCP, HTTP, the CLI, TypeScript or Python.

memory

Keep the reasons

Decisions, rejected ideas, bugs and open questions live next to the code they explain.

decisions

No silent contradictions

Hunch refuses two current decisions that conflict. Replace one, and the history shows what changed.

access

Each agent sees its share

Pick which agents can open each space, down to single records. Add a key per agent to sign what it writes.

actions

Know what’s done and what’s owed

See finished work, open commitments and their sources in one place. Work nobody verified stays marked unverified.

control

You decide what’s a rule

Record how you, your team or your company does things. It stays advice until you trust a rule and turn on strict mode.

git

Your record, your repo

Readable files in Git. Review a change, revert it, rebuild the index any time.

how hunch works

Remember. Retrieve. Check.

Hunch gives the agent a short brief and a way to check its work. The agent still does the work.

01 / remember

Save what you learned.

Capture decisions, corrections and findings, linked to the code and the source they came from.

02 / retrieve

Bring it back at the right moment.

When a task starts, Hunch picks the memory that matters and delivers it through MCP or your assistant’s hooks. What arrives depends on the assistant.

03 / check

Check, then show the receipt.

Hunch runs the rules you trust and writes a task report: what it delivered, what the agent used, what actually passed.

see the evidence

Don’t trust the agent. Read the receipt.

Every task report keeps three things apart that most tools blur together.

npm i -g @davesheffer/hunch
deliveredwhat Hunch handed the agent
reportedwhat the agent says it used
checkedwhat rules and commands actually ran

A green test alone doesn’t prove Hunch helped. No evidence, no claim — it stays unverified. Read a task report →

inside the coding tools

Know what a change breaks before you make it.

For code, Hunch connects the reasons to the dependencies and to the rules your team chose to protect.

why

Find the reason

Decisions, rejected ideas and past failures behind any file or function.

blast radius

See what depends on it

Follow the code’s real connections to see what a change can reach.

project DNA

Learn the house style

See the terms and habits Hunch observed in this repo. Advice only.

compare

Compare options

Line up candidate branches against what you already decided.

conform

Check the code still keeps its promises

For example: does the payment path still check authorization?

change gate

Guard the rules you trust

Flags changes that break a trusted rule. Strict mode blocks only where that rule and your assistant allow it.

Want the technical details? Read how Hunch works →

get started

Start in one repository.

Install Hunch

Install the CLI, then run hunch init in the project you want it to remember.

Teach it your history

Optional: backfill the last few months of Git, then review what it captured.

Ask why

Reload your assistant and ask why a file is built this way. Using Codex? Trust the hooks in /hooks first.

Claude CodeCursorVS CodeWindsurfCodexAntigravityAGENTS.md
  ~/your-repo
# 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 advises by default — nothing blocks until you say so. To update, run hunch update, reconnect your assistant, and in Codex review changed commands in /hooks.

Claude Code? Install as a plugin instead:
/plugin marketplace add davesheffer/hunch → /plugin install hunch@hunch
Open VSX

One record, from customer issue to verified fix.

That’s the goal: a customer reports a problem, an agent fixes the code, another confirms it’s done. Shared state ships today; the Sofia pilot is testing the full handoff.

Get started Read the docs