how it works
Six things happen between the first mistake and the second.
A session is read, a decision is recorded with its reason, the next plan is checked against it, and what happened afterwards goes back on the record.
stage 01
the problem
The second time costs what the first one cost.
A session ends and what it taught ends with it. The next agent starts from nothing and asks the same question again.
recall counts what it did not know. Analytics keeps the unanswered questions with a teach it action on each one, and Home puts what needs a person in one strip.
stage 02
capture
It happens once. You do nothing.
Sessions are read in the background, by your model, on the endpoint you configure.
- Sessions shows what shipped, by project and state, and a failed slice can be retried.
- Extraction and the models that run it are set by you, under Settings: Models, Extraction, Limits.
- Secrets are redacted before storage: key shapes, labelled tokens, connection strings, card numbers.
stage 03
record
It is written down, with the reason.
One record, one URL: the atomic line, the text, because, holds when, and the cue.
- What was actually said is kept beside it, and saidBy marks it as a person's words or an agent's reply.
- History and chain hold what it supersedes, what it led to, what precedes it and what contradicts it.
- On the record itself: ask, edit, move, link, confirm, supersede.
stage 04
the catch
It is checked before the next act.
Plan check runs from any MCP client before the tool call, not after, and from the playground under Developers.
The playground runs recall, deep, brief, plan_check and read against your own spaces, so you can read what an agent would be handed before you wire one up.
stage 05
outcome
It learns from what happened.
An outcome is one click on the record page: held, reversed or regret.
- A reversed decision is marked in every brief after it.
- Coming, not live yet: a nightly pass that proposes principles from what held, for a person to accept or reject.
- Conflicts lists every contradiction: keep left, keep right, both with a condition, or not a contradiction. A person picks; a model never erases either side.
stage 06
control
It is yours.
Spaces say what belongs where. An agent key says how far an agent may go.
- A space page carries what belongs here with a router check, who reaches it, and export as Markdown or JSON.
- Agents have their own page, with edit and key rotation. A key carries a ceiling, and read-only keys exist.
- Type
<recall-off>in a message and the whole session is never shipped.
where it runs
Four ways in, once you have access.
In the app, Developers holds Connect, the playground, keys with a ceiling and the reference.
Claude Code
Set up from Connect in the app. The client ships commit messages as their own slices, subject and body, never the diff.
Any MCP client
Cursor, Codex, Claude, ChatGPT and any other client that speaks MCP use one endpoint: app.recall.hiretrainva.com/mcp.
Keyed MCP
The same endpoint with a key on it. Keys are issued under Developers with a ceiling, and read-only keys exist.
Ship text
Connect also hands you the ship text to paste into a client, with the tool reference next to it.
chat and extraction run on the endpoint you set, including a local one.
early access
Connect, once you have access.
Early access is by invitation while we load-test capacity and cost. With access, every MCP client uses one endpoint.
https://app.recall.hiretrainva.com/mcp