project memory for AI coding assistants

Your AI can read the code.
Hunch explains why it is that way.

Hunch gives Claude, Cursor, Codex, and other coding assistants the decisions, bug fixes, and warnings your team learned before they change your code.

It works with the AI tools you already use. It is not another model, and it does not send your project memory to a Hunch service.

Get started See how it works

a familiar story

The bug was fixed. The reason was not.

A non-obvious safeguard survives in the code. The story behind it slowly disappears.

Monday

A team solves a painful logout bug.

They move sessions server-side so a leaked token can be revoked immediately. The choice adds complexity, but it closes the hole.

Months later

The code remains. The context fades.

The incident is buried in an old pull request. Two people have changed teams. The unusual session flow now looks like needless machinery.

Next session

An AI assistant offers to “simplify” it.

The change is tidy and locally correct. It also reopens the exact failure the team already paid to understand.

With Hunch

The reason arrives before the edit.

The assistant sees what was chosen, what was rejected, and which bug the choice prevents. It takes a better path without asking the team to retell the story.

what Hunch does

It gives every coding assistant the same project memory.

Hunch saves what your team learns, brings it back before an edit, and warns when a change could repeat an old mistake.

01 / save

Keep what the team learned.

Decisions, test failures, and your corrections become project memory instead of disappearing into old chats and pull requests.

02 / remind

Explain before changing code.

The assistant sees why a file is built this way, what depends on it, and what broke there before.

03 / warn

Catch repeated mistakes.

If a change goes against a trusted decision or reopens a known bug, Hunch explains the conflict. Blocking is optional.

what improved in v1.19

Find the likely code sooner. Inspect less.

In a 12-problem test on unfamiliar code, Hunch found the changed piece of code in 6 cases instead of 3. In a separate test, it kept the same five successful finds while asking developers to inspect 41.9% less code.

npm i -g @davesheffer/hunch

3/12 → 6/12problems where the changed code was found

8/12 → 10/12problems where the correct file was found

18.9 → 11pieces of code to inspect on average

These were small, controlled tests—not a promise that Hunch is twice as accurate everywhere. When it only knows where to investigate, Hunch says so instead of pretending it knows the exact fix.

what happens behind the scenes

Hunch stays under your control and shows its work.

Your project memory is made of readable files you can review and undo. Hunch connects each saved reason to the code it affects and includes the source with every answer.

saved with Git

Memory you can inspect

Decisions are saved as plain files. Your team can review, compare, and undo them just like code.

code connections

Knows what a change may affect

Hunch follows the connections between files and functions so the right reason appears in the right place.

your AI tools

One memory across assistants

Claude Code, Cursor, VS Code, Windsurf, Codex, and others can all use the same project memory.

clear rules

Checks the same rule every time

For rules your team chooses to trust, Hunch checks the code directly instead of asking another AI to guess.

source included

Shows where answers came from

Every answer points back to the decision, bug, commit, or correction that supports it.

no Hunch cloud

Your memory stays yours

Hunch does not require a hosted account. Keep memory in your project or in a private Git repository your team controls.

Want the technical details? Read how Hunch works →

get started

Give your next AI session a memory.

Install Hunch

One command reads your project and connects the coding assistants you already use.

Add recent history

hunch backfill --since 90d finds useful decisions and lessons in the last 90 days of commits.

Ask a real question

Try “why is the session module built this way?” Your assistant answers from the team's history, with evidence.

Claude CodeCursorVS CodeWindsurfCodexAntigravityAGENTS.md

~/your-repo

# install from npm — Node 22.13+
$ npm install -g @davesheffer/hunch

# connect Hunch to this project and your assistants
$ cd your-repo && hunch init

# optionally learn from the last 90 days
$ hunch backfill --since 90d

# ask what a file is for
$ hunch why src/auth/session.ts

npm i -g @davesheffer/hunch && hunch init

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

Make your codebase remember why.

Keep the decisions your team already made available to every person and every assistant that comes next.

Get started Read the docs