decision memory: the memory of record for teams and their agents
Your agent already made this mistake once.
No agent repeats a mistake a person already corrected, and you can show who decided what and when.
- Sessions are read in the background.
- Decisions, reasons, exceptions and outcomes become dated records that nothing overwrites.
- When two records disagree, a person settles it, never a model.
- Every agent is briefed before it acts and has its plan checked against what was decided.
picking up the webhook handler work
- preferenceDana prefers small pull requests
- eventStaging double-charged
- personDana, engineering
works withClaude CodeCursorCodexClaudeChatGPTany MCP client
sound familiar
Sound familiar?
If one of these is yours, the rest of this page is for you.
- I told it this last week.
- The CLAUDE.md is 600 lines and nobody reads it, including the agent.
- Which of these notes is still true?
- Did I say that, or did the agent decide it?
- I don't want my manager reading my agent's notes.
- Every new session starts from zero.
one correction, start to finish
Say it once. The next plan already knows.
One sentence said once, followed from the conversation it came from to the plan it corrects. Illustrative; the company is made up.
"Don't offer Acme the standard package. They require enterprise billing because of procurement."
The raw conversation exists. Nothing else has happened yet.
Your model reads the session and writes one dated record: the line, the reason, when it holds, and who said it, with the quote.
The record is linked to the company, the plan, the teams and the contract it touches. Ask about any of them and it is one hop away.
- the conversationclosed
- the agentreplaced
- the sessionended
- the weekpassed
The conversation closes, the agent is swapped, the session ends, weeks pass. The record stays, with its history.
Three weeks later a different agent asks before it acts. The decision comes back with its reason, and the plan changes.
what only recall does
What only recall does.
Other tools remember. These three mechanics are the part they do not do.
01
A person settles contradictions.
When two notes disagree, you pick; a model never erases one. Other tools resolve it on their own, by model or by recency.
02
Decisions carry outcomes.
Held, reversed or regretted, and reversed ones are flagged in every later brief.
03
Grants with agent ceilings.
Each agent can read and write only what you allow. No agent can write a rule that binds a person.
Also true.
Useful, and not unique to us.
- Works with Claude Code, Cursor, Codex, Claude, ChatGPT and any MCP client.
- Plan check. The agent asks before the tool call, not after.
- Who said it. Every record says whether a person or the agent said it, with the quote it rests on.
- History kept. A changed decision is superseded and linked, not overwritten.
- Personal stays personal. Nobody else reads it: not managers, not admins, not other people's agents.
- Forget on request. Only when a person asks. It can be switched off per person and respects legal hold.
- Off switch. Type
<recall-off>and that session is never captured. - Export. Any space, as Markdown or JSON.
one session, from inside
What your agent gets back, prompt by prompt.
A brief when the session starts. On each prompt, a close match comes back and the agent answers with it; nothing close, and it answers alone. What the session taught is filed at the end. Illustrative.
Why something was chosen. It carries the reason, the condition under which it holds, and who said it. A new decision supersedes the earlier one; the earlier one stays in the history.
How someone likes a thing done. One type with a polarity and an intensity, so a mild lean and a hard rule are the same shape.
What failed and what to do instead. Checked against a plan before acting, so the same mistake is not paid for twice.
Who someone is: a name and a card. Never a fact about them; facts link to the person instead.
What happened and when, with its outcome. A decision leads to an event, and the outcome feeds back into the decision.
What a record is about. Subjects live here so they never become types, and a topic gathers every record filed under it.
each person
What changes for each person.
Three roles, one set of records.
the lead who sets rules
Set a rule once.
- Write it with the reason and when it holds. It stands until you change it.
- Contradictions come to you. You pick; neither note is erased.
- Mark a decision held, reversed or regretted. Reversed ones are flagged in every later brief.
- Show who decided what, and when, from the record's history.
the teammate
Stop re-teaching the agent.
- Your agent starts briefed with what the team already decided.
- Your personal space is yours: not your manager, not an admin, not another person's agent.
- Type
<recall-off>and that session is never captured. - Export any space you can read.
the agent
Asks before it acts.
- Gets a brief at the start of every session.
- Runs plan check before the tool call, not after.
- Reads and writes only what its grant allows, under a ceiling a person set.
- Its words are marked as the agent's, beside the quote they rest on.
permissions
Each agent reaches only what you allow.
A record lives in one space. Who reaches the space decides who can know it. An agent reaches no further than the ceiling on its key.
- personalone per person, nobody else reads it
- departmentreached through grants
- sharedthe whole workspace
- agent keya ceiling below its owner
through time
A decision changes. Its history stays.
A changed decision supersedes the old one. The old one keeps its place, its date and its reason, and a brief can still answer as of any day.
Standard delivery is 14 working days
Standard delivery is 10 working days
changed by a person · because: annual delivery review · supersedes the August decision · outcome: heldheld reversed regret reversed ones are flagged in every later briefconflicts two records that disagree wait side by side until a person settles themproof
What is measured, and what is not yet.
One public benchmark, with what it does not show. The test that matters most has not run yet.
LoCoMo · long conversations
On 1,540 questions, Zep-style grader, production service. It measures retrieval, whether the right memory comes back. It does not measure the three mechanics above.
All three graders and the caveatsrepeat rate · corrected mistakes
The counted test: how often an agent repeats a mistake a person already corrected, with a CLAUDE.md file, with claude-mem and with recall. It is coming. No numbers until it runs.
How we report numbersquestions
Questions.
01Can a record be overwritten or deleted?
Not silently. A changed decision supersedes the old one, and the old one stays linked in its history with its date and reason. Every change is audited. A record is forgotten only when a person asks.
02I already have a CLAUDE.md. Why add this?
Keep it. A rules file is good at short, stable rules. It grows, goes stale and carries no reasons, conditions or outcomes. recall vs a CLAUDE.md file.
03How is this different from claude-mem?
claude-mem compresses one person's coding sessions into memory on their machine. recall is for a team: a person settles contradictions, decisions carry outcomes, each agent has a ceiling. recall vs claude-mem.
04Can my agent change a decision?
An agent key has a ceiling. It may recall and record within its grants. It can never write a standing rule that binds a person, and what it records stays marked as the agent's words.
05What does a teammate's agent see?
What its person can reach, and no further than its key allows. A personal space is read by nobody else: not managers, not admins, not other people's agents.
06Which models does it use?
Yours. Chat and extraction run on the endpoint you configure under Settings. Your data is never used to train anything.
07What is the name?
recall is a working name, and it will be renamed. The endpoint, your data and your export do not change when it does.
early access
Correct it once. Not every week.
Early access is by invitation while we load-test capacity and cost. With access: one endpoint, one key, and the next plan is checked against what your team already decided.
https://app.recall.hiretrainva.com/mcp