> ## Documentation Index
> Fetch the complete documentation index at: https://slackhive.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Memory

> Persistent facts the agent learns from conversations, always visible on every turn.

Memory is one of the four layers of an agent's brain - see [Concepts](/agents/concepts) for how it fits with the system prompt, skills, and wiki.

Every memory is **inlined into the system prompt on every turn**. There is no retrieval step, no command to trigger - the agent always sees what it has learned.

## Where memory lives in the UI

Open any agent → **Instructions** tab → **Memory** sub-tab.

* A list of saved memories grouped by type
* Content and description for each memory
* A delete button on each memory

## How memories get saved

When something in a conversation is worth keeping across sessions, the agent asks the user first - *"Should I remember that you prefer BigQuery over Redshift?"* - and only writes the memory after you confirm.

Ask-before-saving keeps memory clean and prevents noise.

## Memory types

| Type        | Purpose                                      | Example                                             |
| ----------- | -------------------------------------------- | --------------------------------------------------- |
| `feedback`  | Behavioral corrections, validated approaches | "Don't mock the database in integration tests"      |
| `user`      | Facts about people the agent works with      | "Kai leads the data team, prefers concise answers"  |
| `project`   | Ongoing work context, decisions, deadlines   | "Merge freeze starts 2026-03-05 for mobile release" |
| `reference` | Pointers to external systems                 | "Pipeline bugs tracked in Linear project INGEST"    |

## Memory file format

```markdown theme={null}
---
name: kai_team_lead
description: Kai leads the data team and prefers concise answers
type: user
---

Kai is the lead of the data team. They prefer concise, direct answers
without lengthy preamble. They use BigQuery, not Redshift.
```

| Field         | Required | Notes                                             |
| ------------- | -------- | ------------------------------------------------- |
| `name`        | Yes      | Snake\_case unique key per agent                  |
| `type`        | Yes      | One of `user`, `feedback`, `project`, `reference` |
| `description` | No       | One-line summary shown in the Memory tab          |

## Size budget

Total inlined memory has a cap (`MAX_INLINED_MEMORY_BYTES` in the runner). If an agent accumulates more than the budget allows, the oldest entries overflow and are dropped with a warning. Prune stale memories from the Memory tab.

## Updating and deleting

* **Update** - the agent overwrites the file with the same `name`. The DB uses `(agent_id, name)` as the unique key, so there are no duplicates.
* **Delete** - click × on the memory card. Removed from the database and not reloaded on next start.

Memories persist until you delete them. There is no TTL.

## Next steps

<CardGroup cols={2}>
  <Card title="Concepts" icon="diagram-project" href="/agents/concepts">
    How memory fits with system prompt, skills, and wiki.
  </Card>

  <Card title="Coach" icon="comments" href="/agents/coach">
    Turn conversation insights into skills and wiki pages.
  </Card>
</CardGroup>