The memory layer for AI coding agents
Your agent forgets your project every new chat session.
Every new session starts with you re-explaining what you built, what you decided, and what's next. Checkpoint remembers all of it — and hands it to your agent the moment a session starts, so it just continues where you left off.
Current phase
2/4
Pipeline board · tasks done
Open bugs
2
Sorted worst-first
Waiting on you
1
Open questions from your agent
Last activity
Jul 12
State digest updated
State digest
what your agent reads at session start · updated Jul 12, 4:42 PM
Where we left off
2026-07-12: Drag-and-drop reorder works; left off wiring optimistic updates in src/app/board/actions.ts — server action exists, UI not calling it.
Current phase: Phase 3 — Pipeline board
Open bugs
- [high] Deal amounts lose cents on save (probably integer column)
Recent decisions
- Postgres over SQLite — hosted from day one, avoids a migration later
Click through the tabs — this is the dashboard your agent keeps current. You read, it writes.
~530
tokens to catch your agent up
re-explaining from Notion costs ~3,800
7×
cheaper than making the agent re-read everything
measured on a real project, not a benchmark
0
tool calls to get caught up
a hook injects state before you type a word
If you ship through an agent, you've lived this loop
01
The chat fills up mid-feature
The context window runs out before the work does. Halfway through a feature, the session that knew everything is gone.
02
You paste summaries back into the chat
You re-explain from memory — or paste a summary of the dead session into the new one — while the agent crawls the codebase to catch up, burning the context you just freed. It still misreads where things stand.
03
The project drifts
Features get marked done that aren't, settled decisions get relitigated, finished work gets redone. You spend sessions steering instead of shipping.
Checkpoint breaks the loop.
Project state lives outside the chat window. Sessions fill up, end, and restart — the memory doesn't.
What your agent gets
- A ~530-token state digest at session start — no codebase crawl to rebuild context
- The exact line where the last session left off
- Decisions with their why, so it never relitigates them — corrected by superseding, never silently deleted
What you get
- New sessions that just continue — nothing to re-explain
- Every session captured — even the ones that crash or forget to log
- Every build entry anchored to its exact git commit
A week of building, with a memory
Not features in the abstract — the moments where Checkpoint catches what every agent-built project drops.
Day 1
You start with an idea, not a repo
You and your agent sketch the product map together — features, decisions, and a PRD. Every session after this knows the destination.
# Recipe App — project created
## Phases
[>] Foundations · [ ] Recipes · [ ] Meal planning · [ ] Sharing
PRD saved · 14 tasks seeded
Day 2
A decision gets made — and stays made
You pick your stack and the agent logs it with the why. No future session relitigates it, re-argues it, or quietly swaps it out.
decision — logged with the why
Supabase over Firebase
Firebase auth fought me for a day; Supabase worked on the first try. Settled — don't revisit.
Day 5
The chat fills up mid-feature
The context window dies before the work does. The session reconciles on the way out; the next chat picks up mid-thought two minutes later.
9:12 PM · chat full — session reconciled
Left off: image upload saves to storage; serving URLs next.
9:14 PM · new chat
Picking up from the digest — wiring the serving URLs now.
Day 12
You look up and see a product
Phases march from planned to done, and every logged moment carries the commit it was true at — a timeline of working states you can always step back to.
Foundations ✓ · Recipes ✓ · Meal planning ← current
a3f19c2Image uploads working end-to-end
8d03f7aRecipe pages live with search
b41c9e2Auth + profiles shipped
Every vignette is the real product — the same digest, badges, and restore button you'll see in your own dashboard.
A vocabulary for how projects actually move
Not freeform pages — a small set of verbs your agent calls over MCP as it works. Structure in, dense digest out.
get_project_state
The digest, on demand
log_build
What just got done
log_decision
Settled choices, with the why
add_idea
Park it, don't chase it
add_bug
File it the moment it surfaces
add_question
Flag what needs a human
update_phase
Keep the roadmap true
reconcile_session
Close the loop at session end
The eight above are the core loop, with feature-blueprint and setup verbs alongside them, including full-text search, project docs, and bug & question tracking. Full verb reference. Works with Claude Code, Codex, Cursor, and any MCP-capable agent.
Set up in under ten minutes
Connect your tool
Choose Claude Code, Codex, or Cursor and get the exact connection flow for that tool. Approve the login in your browser.
Wire up a repo
Paste one tailored snippet — hooks for Claude Code or Codex, or an always-on Cursor rule — so every session loads and saves memory.
Just work
Your agent creates the project, logs as it builds, and reconciles at session end. Next session starts already caught up.
Your current state is always free. Your history is Pro.
No metered calls, no token counting — your agent logs as much as it wants on either plan.
Free
$0 forever
Everything, for one project.
- 1 project
- Every feature and MCP verb
- PRD, phases, decisions & open items kept forever
- 30 days of build-log history
Pro
Recommended$10 per month · or $96/yr (2 months free)
For founders running everything through agents.
- Unlimited projects
- Unlimited history — sessions and build log, forever
- Full-history search
- Priority support
Nothing is ever deleted on the free plan — upgrading unlocks your full history instantly.
Questions, answered
Which agents does it work with?
Claude Code, Codex, Cursor, and any MCP-capable agent. Onboarding asks which tool you use and gives you its exact connection and project-memory setup.
How is this different from keeping notes in Notion or a wiki?
Nobody maintains a wiki mid-flow — it goes stale the week you're busiest. Checkpoint is written by the agent as part of its session protocol, so it's always current. And it's state-first: catching up costs ~530 tokens instead of re-reading pages that cost ~3,800.
Isn't a repo instruction file enough?
CLAUDE.md, AGENTS.md, and Cursor rules hold standing instructions — conventions, commands, and style. Checkpoint holds the state that changes every session: the current phase, open bugs, what happened last time, and why decisions were made. The setup connects the two.
What happens when a chat fills up mid-feature?
End the session — the agent reconciles its progress into Checkpoint, including an exact “left off” line. Open a fresh chat and the session-start hook injects the digest, so the new session continues mid-thought instead of starting an investigation.
What if my agent forgets to log — or the session just dies?
That's what auto-capture is for. Claude Code and Codex use a session-end hook that sends the session to Checkpoint, which distills what happened into the log automatically. If the agent already reconciled, capture skips itself. Only the distilled entry is stored — never the raw transcript. Cursor doesn't expose its transcript today, so its always-on project rule makes reconciliation the required path.
What's actually free? Can I cancel Pro anytime?
Free is the whole product for one project — every verb works, so your agent never hits a paywall mid-session. Free keeps 30 days of build-log history; older entries are hidden, never deleted. Pro is cancel-anytime through Stripe, and if you downgrade you keep everything — the deep history just locks again until you return.
Who can see my project data?
Only you. Your projects live in managed Postgres (Supabase), scoped to your account and backed up nightly. We never train models on your content and never sell data. Payments are handled by Stripe — card details never touch our servers. Delete a project and it's gone — and you can export everything as a single markdown file anytime, so there's no lock-in.
Stop pasting summaries of your chat back into itself.
Connect an agent, run one session, and open the next one already caught up — nothing to re-explain.