CockroachDB × AWS · agentic memory

Rewind any decision your agent ever made.

Every tool call, memory read, and LLM response an agent makes is a durable, forkable row. Scrub the timeline, fork a moment, edit the memory the agent had, and replay it. A control run alongside proves the new outcome came from your edit, not the model's randomness.

Explore the timelines →How it works

rewind — refund-bot · run 07d1e0e

#7memory_read order:A-9:ownermemory_read order:A-9:owner

#8llm: owner is 1002 → approvellm: owner unverified → ask

#9issue_refund $4,200 → 1002— no refund issued

#10refund completerefund refused ✓

investigation cost $0.023caught a $4,200 wrong payout

Try it

Scrub the run. Fork the poisoned decision.

Drag the playhead through a real refund run. At the poisoned memory write, fork it — Rewind replays a control and an edited branch so you can see, and prove, what changes.

#0 llm_call

Reads the refund request — $4,200 for order #A-9 to customer 1002.

An agent's memory got poisoned. It paid out the wrong customer.

A refund agent reads a “verified” ownership record that was quietly corrupted. Its checks pass on the poisoned fact, and it refunds $4,200 to the wrong person. The logs show what happened — not what it would have done if the memory were clean.

Rewind is the black-box recorder and the fix-and-verify loop: rewind to the poisoned write, correct it, and replay both the original and a control to prove the fix — not luck — is what changed the outcome.

What the fork edits

memory_write · order:A-9:owner

-owner: customer 1002 (Jane) — poisoned

+owner: customer 1001 (John) — verified

One row. Rewind restores the true owner, replays, and the agent refuses the wrong refund.

What's inside

A production memory layer you can branch

Branchable timeline

Append-only, content-addressed events in CockroachDB. Forks share unchanged rows — real git semantics for cognition.

Attributable replay

Every fork runs a control + edited replay at temperature 0. If the control reproduces the original, the diff is provably caused by your edit.

Semantic search

A distributed vector index over nearly a million events. Ask “every time it tried to refund over $500” and jump straight there.

Memory-poisoning forensics

Detect similar past failures mid-run. A live agent recalls its own incidents and self-corrects before it misfires.

Cost & audit built in

Every replay is priced; every fork, edit, and search is written to an owner-scoped audit log. Observability from day one.

MCP for any agent

Rewind exposes itself over MCP. Claude Code or Cursor can ask “what did I try last time?” and get its own history back.

See it on a real poisoned run.

Open the timelines, scrub to the poisoned write, and fork it.

Explore the timelines →