Fix it once.
Never twice.
Your coding agent forgets everything between sessions. Ghost Memory reads your history, extracts what you already taught it, and feeds it back at the right moment.
Prefer pip? pip install git+https://github.com/mathischarlesgauthier/ghost-memory
Runs on your existing history. Nothing leaves your machine.
Every correction you make: paid once, reused forever.
Your agent doesn't need a bigger brain. It needs a memory.
The knowledge isn't in the docs. It's in the 7 times it broke before it worked.
Nobody writes skills. That's the whole problem — and why Ghost Memory writes them for you.
Every correction you make: paid once, reused forever.
Your agent doesn't need a bigger brain. It needs a memory.
The knowledge isn't in the docs. It's in the 7 times it broke before it worked.
Nobody writes skills. That's the whole problem — and why Ghost Memory writes them for you.
Your agent doesn't need a bigger brain. It needs a memory.
Stop repeating yourself
Your agent stops rediscovering what you already taught it.
Your codebase, not generic advice
Ghost Memory only learns from your sessions, your repo, your conventions.
Zero effort
No files to write. No skills to hunt for. No config.
Paid once, reused forever
Every correction you make becomes capital. The knowledge isn't in the docs — it's in the 7 times it broke before it worked.
Local-first
Runs on your existing history, on your machine. Nothing is published without you seeing the diff first.
Three commands. That's the whole workflow.
One command. Zero config. And it runs on the history you already have — you see the result in 90 seconds.
read your past sessions
distill them into skills
wire them into your agent
That's it. Then you just work.
You explain that migrations on your ledger table have to go through batch_alter_table. It gets it. It works.
New session. It doesn't know. Seven attempts. Forty minutes.
Your new hire hits the same wall.
You've paid for that lesson three times.
Your agent's knowledge is worthless. Your scars aren't.
What Claude gets right on the first try.
It's in the model. Everyone has it. Any prompt regenerates it.
The 7 times it broke before it worked.
The 12 times you said "no, not like that."
That doesn't exist anywhere except in your history. By definition, no model can guess it — if it could, you wouldn't have had to correct it.
Ghost Memory reads your scars.
You already paid for your scars.
You haven't paid for theirs.
Your own history only gives you back what you already learned the hard way. Useful — you never pay for it twice. But you already took the hit.
The wall you'll hit next Thursday, someone already hit it. Their scar is distilled, measured, and waiting.
You cannot build that alone. No matter how good you are.
Free
$0 · foreverYour own memory.
- ✓Reads your entire session history — 400 sessions, all of it
- ✓Distills skills automatically — you write nothing
- ✓Refuses to distill noise — candidates get skipped, on purpose
- ✓Injects the right skill at the right moment — you search for nothing
- ✓Unlimited. No request caps. No quota. No seat limit.
- ✓Runs on your own API key — your code never leaves your machine
- ✓Private by default — nothing published without you seeing the diff
- ✓Open source — read every line before you run it
Free because it costs us nothing. Your machine, your key. Not generosity — honesty about our costs.
Pro
$29 / monthEveryone else's memory.
Everything in Free, plus:
- ✓Every skill ranked by measured effect on real code — not by download count
- ✓Ranked for YOUR stack — the same skill can help you and hurt someone else. We know which
- ✓Skills for mistakes you haven't made yet — the ones that cost someone else a weekend
- ✓Skills that hurt get demoted automatically — the index cleans itself
- ✓Unlimited replay validation — we rerun your own past tasks with and without a skill, and show you the delta
- ✓Continuous holdout — we withhold skills on a slice of your sessions on purpose, so the number stays honest
- ✓Method visible on every number — no black box
- ✓Premium skills — distilled from hundreds of hours in production, not regenerable by any prompt
- ✓Sync across machines
- ✓Publish and sell your own — you keep 80%
- ✓Public profile — installs and measured effect. It's a CV.
Cheaper than one failed session a month.
There are 670,000 agent skills out there. They're ranked by download count — a number you can buy, and that tells you nothing.
A benchmark scored 47,150 of them. Average: 6.2 out of 12.
Most do nothing. Some make your agent worse.
We don't count. We measure.
You're skeptical. Good.
"I already have a CLAUDE.md."
"There are already 670,000 free skills."
"It reads my code?!"
"Another tool to configure."
Lift is the only number that matters.
How much a skill actually improves a task — measured by rerunning it, not guessed from a download count.
670,000 skills out there, ranked by download count — a number you can buy.
We rank by what a skill does to real code.
We don't count. We measure.
Every real number on this site comes from replaying your own past tasks. Until yours are measured, the examples above are just that — examples.
Six months of debugging. None of it saved.
Until now.
Turns your coding history into skills your agent actually uses.
You already wrote your best skill. You just don't know it yet.
Your coding agent has goldfish memory.
Monday you teach it something. Thursday it makes the same mistake.
I built Ghost Memory: it reads your session history, pulls out what you already taught it, and feeds it back automatically.
uv tool install git+https://github.com/mathischarlesgauthier/ghost-memory — runs on your existing history.
I pointed it at my own 412 Claude Code sessions.
47 skill candidates. 312 failure loops. 89 times I had to correct it by hand.
Six months of scars, sitting there, unused.
Nobody writes agent skills because writing one is 2 hours of unpaid work.
That's why 80% of the 670k public skills are slop.
So we stopped asking people to write them. We extract them from what you already did.