Give your agent context that stays true.
Every session starts cold. You re-paste the same context — and half of it is already stale, so your agent acts confidently on last month's decision.
ctx-loop gives you — and your team — living context: small, sourced, expiring cards your agent pulls in fresh, every time.
Be first to run v1. One email when it's ready — no spam.
get "ledger + auth refactor" → 2 fresh · 1 excluded
We use Postgres, not Dynamo, for the ledger — chosen for transactional integrity.
Staging Redis is single-node; load tests flake over 2k rps.
Context is a stream, not code.
A new wave of memory layers helps your agent remember more. That's not the problem.
The problem is that what your team already knows goes wrong: context arrives constantly — a decision, a gotcha, a spec change — and it decays — last month's architecture call is now false. Remembering more of it just means acting on more of the rot.
ctx-loop isn't a bigger memory. It's the maintenance loop that keeps what your team knows true — every card owned, sourced, and expiring.
// capture → approve → serve → decay → re-confirm
Capture
Paste a note; the agent drafts a card — summary, type, tags, TTL, owner.
→Approve
Glance at the draft and its blast-radius, tap keep or reject. Seconds.
→Serve
Your AI agent pulls the freshest relevant cards in-session, over MCP.
→Decay
Past-TTL cards are flagged, dropped from serve, and surfaced to re-confirm.
The agent does the logistics. You make every keep / reject call — in a glance.
Plugs into the agent you already use
MCP server + CLI. Works in Claude Code, Cursor, any MCP client — your agent reads context and captures it for you.
Cards, not a 5,000-line dump
Small, self-contained, sourced. get returns the freshest few relevant cards — not your whole repo re-pasted.
Every card expires
A TTL by type. Stale cards are flagged and never served to your agent— so it stops acting on last month's truth.
Approve in seconds, not minutes
The agent drafts the card, computes blast-radius, and recommends keep or reject. You glance and tap. Target: ≤10s.
Every card cites its source
A provenance link on everything. Your agent can trust it and cite it; you can audit it. No unsourced context, ever.
Starting with your session. Built for your team.
The loop gets its real power when it's shared. One living shelf your team curates together:
One shelf, everyone fresh
A shared context store the team maintains together. When a card expires or is superseded, that decay propagates — every teammate's agent stops serving it, at once.
Every agent, every team, same truth
AI sessions across your team — or across teams — pull from the same fresh, sourced context. No more "whose context repo is right?"
Share exactly as much as you choose
A store can be private, team-shared, or partially shared. Start with your own; share the cards worth sharing. You never migrate everything into one giant org context layer to get value.
v1 runs local-first on your machine. The team layer rolls out to the waitlist cohort next — tell us below how you'd use it.
"Won't it be empty on day one?"
Point ctx-loop at your existing context repo or docs — seed mines them into starter cards. You open to a stocked shelf, not a blank one.
"Another giant context platform to feed?"
No. Cards are small and yours: a local store you own, private by default, shared only when you choose. Vendor-agnostic, and seed stocks it from what you already have.
"What about secrets?"
Secret-shaped input is refused at capture — never stored, never sent to an LLM. A guardrail, not an afterthought.
Be first to run the loop.
We're onboarding the first cohort by hand. Tell us how you'd use it and you'll move up the list.