A coding agent with a memory instead of a context window. It never forgets what you decided, never makes you open a fresh chat to get it thinking straight, and turn 90 is as sharp as turn 9 — on the same frontier models (Anthropic, OpenAI, xAI) as the $20–$200 plans you already know.
$ curl -sSL https://storage.googleapis.com/assertion-downloads/install.sh | sh $ cd your-repo && assertion
One checksum-verified assertion binary in /usr/local/bin; assertion upgrade keeps it current. The Mac app is signed and notarized by Apple — it opens with a double-click.
A real 24-turn build task, run start to finish two ways: once on a normal agent, once on Assertion. Same prompts, same tests passing at the end, same nine questions asked about decisions made hours earlier.
Measured on a real build task (a Python package, models → store → CLI → persistence → refactor → docs), 116–250 model calls per run, the same model in both arms. The one thing memory did worse: the from-memory CHANGELOG had 33 bullets to the transcript's 47 — right and cited, but coarser.
One line in a terminal, or the Mac app. The tools, editor integrations and models you'd expect are there on day one.
On the first run, the agent reads your README, your agent files, your docs and the last 50 commits into the project's memory. Your first question gets a cited answer about your own project — before you've told it anything.
"Which queue did we settle on, and why did we move off SQS?"
— asked hours later, long after that conversation scrolled away
"Redis Streams. SQS dropped a message on restart, so the worker moved over in commit abc1234."
You never see a token bill. Each plan includes a week of usage, shown as a percentage that resets Monday. Because each dollar of usage does about twice the work, the included week goes further than its price suggests.
| Plan | Price | Cursor equivalent | Included |
|---|---|---|---|
| Free | $0 | Hobby · $0 | About 15–20 turns a week. Enough to see it work, not enough to live on. |
| Light | $15/mo | Pro · $20 | A working week for one person. |
| Pro | $45/mo | Pro+ · $60 | 3× Light. All-day use, the larger models when you want them. |
| Heavy | $150/mo | Ultra · $200 | 20× Light. For people who run the agent all day, every day. |
One plan per team space, billed per active member. $30 a standard seat, $90 a premium seat (5× the allowance). Usage is pooled across the team with a per-person ceiling, so one heavy week can't drain everyone else. A team of ten on standard seats is $300 a month; Cursor Teams is $400.
Real compute, metered honestly — a larger model draws the week down faster, and the picker says so where you choose. Memory capture is included.
Go through Assertion's gateway to Anthropic, OpenAI or xAI on our keys — the same providers behind the plans you use today. You never hold or paste a provider key.
Your project's memory: the decisions, their reasons and where they came from — not your repository. It lives in Assertion's cloud, private to you per repository unless you share it.
SOC 2 audit in progress. Independent penetration test completed, with no critical, high or medium findings.
A space is shared only with the people its owner adds, and everything captured there is visible to its members. Owners manage members and seats.
Signed with our Developer ID and notarized by Apple. Apple silicon for now — no Intel build.
Documents written "from memory" list decisions, not every edit. If you want a changelog of every line changed, ask it to read git.
Already on Claude Code, Codex or Cursor? The same memory ships as a plugin for those — install it here. But the 2× is the agent's, not the plugin's.
A bigger window still fills up — and when it does, your agent throws things away to make room. That's the moment it starts re-asking what you already answered. Here there's nothing to throw away, which is why the sixth hour works like the first and costs like it too.
No — a summary loses exactly what you need later. It keeps your decisions, the reasons behind them and the things that bit you, each one you can look up and trace back to where it came from. You can read everything it knows about a project, and correct it when it's wrong.
Not in the alpha. Models are included in the plan and go through our gateway; that's how usage stays a percentage instead of a bill.
The agent tells you plainly, with the reset day and an upgrade link. Nothing is lost; the memory is still there when the week resets.
Yes. A repository's memory is private to you by default; from the app you can share it with teammates, or point a repo at a team space. Team plans bill per active member and pool usage.
Not yet. The installer and the Mac app are macOS on Apple silicon for now.