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

# Memory

> Three-layer memory system that makes agents smarter over time.

# Memory

Miteos agents have a three-layer memory system that enables learning, context retention, and personalization.

## Memory Layers

### Working Memory

**Scope:** Current task only. Cleared after task completion.

Contains the agent's immediate context: what it's working on, what it's found so far, intermediate results. Think of it as the agent's "RAM."

### Session Memory

**Scope:** Per-workspace. Persists across tasks within the same workspace.

Stores context relevant to a project or workflow. If you're researching competitors in a workspace, the agent remembers previous findings when you submit a new related task.

Backed by Redis for fast access.

### Persistent Memory

**Scope:** Per-user. Permanent.

Long-term knowledge about your preferences, patterns, and history:

* Code style preferences
* Communication tone
* Preferred data sources
* Brand voice guidelines
* Timezone and formatting preferences

Backed by PostgreSQL with pgvector for semantic search.

## Managing Memory

### View All Memories

```bash theme={null}
GET /memory
```

Returns memories across all layers with metadata:

```json theme={null}
{
  "memories": [
    {
      "id": "m1",
      "layer": "persistent",
      "key": "code_style",
      "value": "Prefers functional components with hooks. TypeScript strict mode.",
      "source": "Learned from 12 tasks",
      "confidence": 0.95
    }
  ]
}
```

### Update a Memory

```bash theme={null}
PUT /memory/{memory_id}
{"value": "Updated preference text"}
```

### Delete a Memory

```bash theme={null}
DELETE /memory/{memory_id}
```

## How Agents Use Memory

Before acting, agents query relevant memories:

1. **Semantic search** — Find memories related to the current task
2. **Recency weighting** — Recent memories rank higher
3. **Confidence threshold** — Only use memories above 0.7 confidence
4. **Source attribution** — Agent knows where the memory came from

## Learning

Agents learn automatically from:

* **Your corrections** — When you guide or redirect, agents store the preference
* **Successful patterns** — Repeated successful approaches get memorized
* **Explicit settings** — Preferences you set directly (via /preferences endpoint)

## Privacy

* All memory is per-user and encrypted at rest
* Agents cannot access other users' memories
* You can delete any memory at any time (GDPR compliant)
* PII is automatically redacted from agent logs