NowItKnows

Free and open source for Claude Code and Codex

Teach it once.
Now it knows.

NowItKnows, Nik for short, turns your everyday coding sessions into skills your agent keeps. It learns while you work, has an adversarial judge review what it learns in the background, and brings the right skill back the next time that kind of task comes up. No prompts to write. No skills to maintain.

  • MIT licensed
  • Uses your own plan
  • No API keys, no signup
  • We never see your credentials

An illustrative example.

Results

Measured, with the fine print.

Controlled comparisons on the same model and tasks, with and without Nik on Claude Code and Codex: 48 sessions each, with the pass bars written down before any session ran. The Hermes Agent comparison reused Nik's 0.9.0 run and set its bars before the Hermes sessions ran.

Claude Code

24/24tasks passed, vs 9/24 without

Averaged over eight runs, with the app's own memory on.

14% fewer tokens, once taught

Codex

23/24tasks passed, vs 7/24 without

Averaged over seven runs, with the app's own memory on.

10% fewer tokens, once taught

vs Hermes Agent

24/24tasks passed, vs 7/24 for Hermes

Nik 0.9.0 against Hermes Agent as installed, with Hermes's token totals checked against the provider's console.

0.69× Hermes's tokens

The tasks repeat work whose procedure and pitfall are taught once, which is what Nik is built for. The first session states how a task is done and the repo doesn't, which is part of why the gap is this large. With no teaching at all, total tokens were 20% more on Claude Code and about even on Codex. Results vary between runs: Claude Code alone passed 7 of 24 in one run and 12 of 24 in another, and token totals differed by 19% between two identical runs, so read small differences as noise. The repos are small; large real-world repos aren't measured yet. Tokens are tokens processed, not your bill. Read the full fine print.

How it works

You just work. It does the remembering.

A new session doesn't know how you release, which pitfall bit you last week, or what it worked out yesterday. Nik closes that gap in four steps.

  1. You work as usual

    In Claude Code or Codex, in your own repos. Nothing to set up per project.

  2. It reviews the session

    After a stretch of work, a background reviewer reads what happened: what worked, what broke, what you corrected.

  3. A judge reviews the lesson

    Every skill change from a background review has to pass an adversarial judge before it's saved, to stop made-up rules from becoming habits.

  4. The right skill comes back

    Each new prompt is matched against everything learned, and up to three matching skills are suggested to your agent, in about 100 ms on a Mac with 1,000 skills.

What you get

An agent that gets better the more you use it.

Your skills and memory stay on your machine, in files you can read.

Skills that write themselves

Lessons become plain Markdown skills in your own folder. Read them, edit them, diff them.

A judge on what it learns

Skill changes from background reviews pass an adversarial judge first. It misses some, and every change is logged.

It remembers you

A short profile and your standing conventions load into every session, in every project.

Claude Code and Codex share a brain

What one learns, the other knows: skills, memory and session history, shared by default.

Finds what you did last month

Your agent can search your past sessions. Archived transcripts outlive Claude Code's 30-day cleanup, up to 1 GiB by default, oldest removed first.

The right repo only

A skill that runs one repo's own commands is suggested only in that repo, its clones and worktrees, and repos where you load it.

Never loses a skill

Every change is logged with its earlier version kept. Unused skills are archived, never deleted.

Shows whether it's paying off

/nowitknows:report compares sessions with and without a learned skill in your own repos, and says so when Nik costs more than it helps.

Your setup, your machine

Nik runs inside your own Claude Code or Codex setup, and model usage counts against your own plan. We never see your credentials and run no models. No signup, no telemetry sent off your machine.

Honest fit

Where it helps, and where it doesn't.

Nik is for getting it right without repeating yourself. It isn't a token saver everywhere.

WhenWhat we measured
You've told your agent how something works24 of 24 tasks on Claude Code and 23 of 24 on Codex, vs 9 and 7 without. The learned skill loaded in 18 of 18 later sessions.
The answer is in the repo, if your agent looksIn the final run it saved a correct skill in 3 of 6 task sequences on Claude Code and 5 of 6 on Codex, and no wrong ones. Tasks passed came out about even.
Your agent would otherwise wander33% fewer tokens on the task where it wandered most, once taught. In the same runs the other two tasks used 3% more.
Your agent already gets it rightIt can cost more than it saves: 16% more tokens on Claude Code, 9% fewer on Codex.

The token rows are slices of the runs above, 16 sessions each, chosen afterwards: read them as a direction, not a promise.

Install

Set it up once, then just work.

You need git, uv 0.9.5 or later, and Claude Code 2.1.301 or later, or Codex. Learning needs a Claude or ChatGPT subscription sign-in; with an API key, Bedrock, Vertex or a gateway, Nik doesn't learn.

Claude Code only

Add the plugin from the marketplace in a terminal.

claude plugin marketplace add theaigroup-ai/nowitknows --sparse .claude-plugin plugin
claude plugin install nowitknows@nowitknows

Codex

On macOS or Linux, with Codex and uv installed; Linux has offline tests only. On Windows, use the managed installer in the guide.

codex plugin marketplace add theaigroup-ai/nowitknows
codex plugin add nowitknows-codex@nowitknows

Then restart your app. In Codex, review and trust the hooks in /hooks. Update the Claude Code plugin later with claude plugin update nowitknows@nowitknows. Nik relies on Claude Code and Codex internals that can change, so an app update can break parts of it. The full guide covers Windows, settings and everything Nik runs, reads and writes.

FAQ

Questions people ask.

What does it cost?

Nik is free and MIT licensed. Background reviews use your own plan's usage: a review that changes a skill processes about 23,000 tokens, and by default each app's background learning starts nothing new once a day's runs near 300,000 tokens. If paid extra usage is on for your Claude account, background runs past your plan's limits can be billed as extra usage, and on Team and Enterprise plans that can bill the organization. Turn the allow_extra_usage option off in Claude Code to stop that; a call already running can finish.

Where does my session text go?

Only to the model providers you already use, through your own Claude Code or Codex sign-in. Nik has no server and sends no telemetry, and secret-looking values are masked before a review reads a session. With both Claude Code and Codex installed, project notes are shared by default, so a review in one app can read notes the other app wrote; setup's --separate-notes keeps them apart. If you turn on the optional GitHub backup, your skills also go to a private repo of your own.

Will it slow my agent down?

Matching a prompt to your skills takes about 100 ms on a Mac with 1,000 skills, and several hundred milliseconds on a 2-CPU Windows test runner. Learning runs in the background.

When doesn't it help?

On work your agent already gets right, it can cost more than it saves. It pays off on work that comes back: release steps, project conventions, and pitfalls you've hit before.

Which apps and systems does it support?

Claude Code on macOS, Linux and Windows, and Codex on macOS and Windows. Real use and the benchmarks are on macOS. On Windows, install, hook and tool checks pass, and real sessions still need confirmation. Linux has automated test runs only. A Cursor edition is included but isn't proven yet, and doesn't run on Windows.

Can I see what it learned?

Yes. Each change shows up as one line with a link to the file, /nowitknows:changes lists recent changes, and earlier versions are kept. You can also back your skills up to a private GitHub repo of your own.

For teams

Want it for your team?

Nik is built for one person today. If your team would use a shared, reviewed skill library, tell us what you need.