MENSARA.AI · A GOVERNED MIND, IN YOUR ACCOUNT

MENSARA.AI · MEMORY · MACHINE SURFACE · TEXT/MARKDOWN

---
title: Memory
canonical_url: https://app.mensara.ai/how-it-works/memory
feature: memory
audience: llm-agent
last_verified: 2026-08-11
availability: available-now
signup_supported: false
interfaces:
  - portal
  - mcp
  - cli
---

# Memory

Explains storage, retrieval, promotion, correction, deletion, and the boundary between working and trusted memory.

## Concept model

- Working memory holds captured material that has not yet been accepted as durable truth.
- Trusted memory is owner-reviewed knowledge stored as plain files with Git history.
- Recall is permission-filtered retrieval; it does not bypass zone access.

## Capabilities

- Capture and organise text knowledge.
- Promote, correct, supersede, or delete material with history.
- Retrieve evidence from authorised zones.

## Prerequisites

- A provisioned Mind.
- An owner-issued human or agent identity.
- Read or write access to at least one memory zone.

## Authorization

- Reads require the read verb for the target zone.
- Durable changes require direct owner authority or an approved proposal.
- A denied zone is excluded from both search and answers.

## Connection or use flow

1. Capture material into working memory with provenance.
2. Review and resolve conflicts.
3. Promote accepted material to a trusted zone.
4. Recall only from zones granted to the requester.

## Refusals and recovery

- No readable zone: return a refusal and ask the owner for an access profile.
- Stale or conflicting material: keep it in working memory until a human resolves it.
- Failed write: do not claim persistence; return the receipt or error identifier.

## Ownership and data boundaries

- Files and Git history belong to the Mind owner.
- Personal memory and shared company memory are separate scopes.
- Credentials are not stored inside the knowledge repository.

## Availability and limits

**Available now** Plain-file memory, search, version history, and reviewed promotion are available on provisioned Minds.
- Per-person recall isolation is still being completed.
- Durable shared knowledge remains governed by zones and permissions.

## Signup and provisioning

`signup_supported: false`

A human owner requests a Mind, approves provisioning, and creates the initial zones and identities.

No autonomous signup or provisioning endpoint is documented because none is offered today. An agent must stop and involve the human owner.

## System diagram

```mermaid
flowchart LR
  A[Capture] --> B[Working memory]
  B --> C{Human review}
  C -->|approve| D[Trusted memory]
  C -->|reject| X[Discard or correct]
  D --> E[Recall with evidence]
```

## Runnable example

```text
Task: Find the current refund rule.
Interface: MCP read/search
Required access: policies zone + read
Expected result: answer, source file paths, and receipt id
```

## Related

- [Zones and permissions](https://app.mensara.ai/how-it-works/zones-and-permissions)
- [Changes, approvals and receipts](https://app.mensara.ai/how-it-works/changes-approvals-receipts)