# Keep what is true. Send only what matters.

From working memory to task-ready context. Unforgetter sits beneath the agents you already use. It captures what each session learns, compiles an approved working truth, serves each task the scoped slice it needs and controls what can become memory afterwards.

- Managed cloud layer
- Vendor-neutral MCP
- Private alpha

Canonical: https://unforgetter.com/how-it-works

## A context layer, not another agent platform.

Your sources remain the record of the work. Your agent still performs the task. Unforgetter maintains and compiles the context between them, and never executes anything on your behalf.

Knowledge plane: your sources and the compiled working truth (projects, decisions, preferences, methods, open threads, provenance).

Unforgetter: maintain and compile. Resolve what still applies, enforce scope, assemble task-ready context, accept proposals through a single publisher.

Execution plane: Claude, ChatGPT / Codex, Cursor and Grok Bot. The agent receives permitted context through MCP and does the work in its own runtime.

## What an agent learns is captured as a proposal, not lost in a transcript.

At the end of a session, or when an agent notices a change, it submits a bounded proposal: a decision, a correction, a new open thread or a rejected idea, with the evidence behind it.

- **Input · Session outcomes**: What was decided, what stayed open, what was tried and rejected. Structured, small and attributable to the agent that produced it.
- **Input · Your corrections**: When you correct an agent, that correction is a first-class input. It becomes part of the record rather than a message in one tool's history.
- **Input · Approved sources**: Notes, decision records and project material you explicitly allowlist. Raw inboxes, secrets and private notes are excluded by default.
- **Compatibility · One memory across your agents** (Working now): Claude, ChatGPT / Codex, Cursor and Grok Bot use the same MCP contract. Automatic capture from every client is not claimed.

## History goes in. Working truth comes out.

The compiler reconciles new input against the current record. It resolves what supersedes what, keeps rejected proposals so they are not re-proposed, marks freshness and refuses changes to protected areas.

Memory: Current status, decisions and their reasons, open threads and rejected proposals.

Methods and skills: How you work: conventions, writing rules, review standards. Portable instructions any agent can follow.

Policy and scope: What each agent may see and do. Project, client and data-class boundaries enforced by the server.

Provenance: Which document, which date, which person or agent. Every line can be traced.

Freshness: When the record was last compiled and which facts are provisional, so stale material is marked rather than trusted.

One task bundle: Small enough to fit the task; Scoped to what this agent may see; Marked with provenance and freshness; Pointers for deeper retrieval instead of a vault dump.

- **Supersession, not deletion**: A newer decision replaces an older one in the compiled state. The older one stays in history with the date it stopped applying.
- **Rejected stays rejected**: Ideas you turned down are recorded as rejected, so an agent does not rediscover them as fresh suggestions next week.
- **Protected targets**: Goals, policies, permissions and identities cannot be changed by an agent proposal. They only change when you change them.
- **Single publisher**: One consolidation process writes the canonical record. Agents read, search and propose. Nothing else writes.

## The useful context is defined by what stays out.

For the task used on the homepage, "Extend the booking flow so a late cancellation is charged against the right date.", the compiler includes current, authorised material and deliberately excludes stale or unrelated material.

Included in the task bundle

- Decision: Fee measures against the original booking date. 6 Aug · approved by you · supersedes the 22 July draft.
- Open thread: The capacity split is unresolved. Do not implement around it until it is decided.
- Method: Money is stored in minor units, never as a float. Your convention, carried into every agent that writes code.
- Scope: Booking product only. Finance and client material are outside this task.
- Provenance: 3 sources, freshest 6 Aug. Every line traces back to a document you can open.

- **Stale · The rejected 22 July fee proposal**: Superseded. Kept for audit, not sent to the agent.
- **History · Four months of session transcripts**: History, not working truth.
- **Scope · Every other project in your workspace**: Out of scope for this task.

A context response is small and bounded: agent identity, the relevant goals and policies, current project facts, open threads, freshness and provenance, plus pointers for deeper retrieval. The whole workspace is never sent to a model.

## Agents can suggest. They do not silently decide.

The record stays current because every session feeds it and nothing overwrites it. Proposals are validated, quarantined when their evidence is untrusted, and published by one controlled process.

Proposal from Codex: add that refunds follow the same original-date rule. Diff: one addition, zero deletions, touches one decision.

Publisher checks: scope allowed, target not protected, no conflict with a newer decision, taint level acceptable.

Published 27 Aug: the updated decision is served to every agent from its next session, with provenance intact.

1. **Observe.** An agent encounters a decision, correction or new open thread during the work.
2. **Propose.** It submits a bounded change with source, scope and the record it would supersede.
3. **Govern.** The publisher checks authority, protected areas, conflicts and taint. Untrusted evidence can raise scrutiny; it can never lower it.
4. **Publish.** The approved state becomes current. The next agent receives the updated version with provenance intact.

## What is live, what is being built, what is not claimed.

| Capability | Status | What you can rely on today |
| --- | --- | --- |
| Versioned Markdown memory | Verified | The canonical record is plain, versioned Markdown you can open anywhere. |
| Same scoped memory across your agents | Compatible | Claude, ChatGPT / Codex, Cursor and Grok Bot connect to the same working memory through MCP. |
| Server-enforced scopes | Working now | Tokens are bound to workspace, trust zone and project catalogue; out-of-scope reads are refused. |
| Provenance and freshness in context | Working now | Every context response carries source references and compile time. |
| Proposal, quarantine, single publisher | Working now | No agent writes canonical memory directly. |
| Revoke, export, deletion | In build | Product functions in the alpha foundation; not yet smoke-tested live. |
| Self-service onboarding, OAuth, teams | In build | Not available yet. Alpha workspaces are connected with you, one client at a time. |

- [Security and ownership](https://unforgetter.com/security)

## Unforgetter does not replace the tools doing the work.

- **Not an agent runtime**: Claude, ChatGPT / Codex, Cursor and Grok Bot and other tools still plan, execute and own their tasks.
- **Not a transcript dump**: History is distilled into working truth and only relevant context is served.
- **Not a proprietary skill format**: Open standards such as MCP and readable Markdown remain the foundation.
- **Not public self-service yet**: Private-alpha workspaces are connected through guided setup while signup and OAuth are in build.

## See what changes when the next agent already knows.

Tell us which agents you use and what you want their shared memory to keep track of.

- [Join the private alpha](https://unforgetter.com/private-alpha)

## Next

- [Company Brain](https://unforgetter.com/company-brain)
- [Security and ownership](https://unforgetter.com/security)
- [MCP reference](https://unforgetter.com/docs/mcp)

---
Status: Private alpha · guided setup. Contact: hello@unforgetter.com.