# Memory

OpenClaw's memory system uses plain Markdown in the agent workspace as the foundational approach. Files serve as the authoritative source rather than RAM-based storage.

## Memory File Structure

The system organizes information across two layers:

### Daily logs (`memory/YYYY-MM-DD.md`)

Append-only daily entries, with today's and yesterday's files loaded at session start.

### Long-term memory (`MEMORY.md`)

Curated persistent information, loaded only in private sessions.

## Writing to Memory

Recommended storage patterns:

- **Decisions, preferences, and durable facts** go to `MEMORY.md`
- **Ephemeral notes and contextual information** in daily logs
- **Explicit requests to remember something** should be written immediately

## Automatic Memory Management

When sessions approach token limits, OpenClaw triggers a silent agentic turn prompting memory consolidation before context compaction occurs.

This flush mechanism can be configured via `agents.defaults.compaction.memoryFlush` settings:

```json5
{
  agents: {
    defaults: {
      compaction: {
        memoryFlush: {
          enabled: true,
          softThresholdTokens: 4000,
          prompt: "...",
          systemPrompt: "..."
        }
      }
    }
  }
}
```

## Search Capabilities

The system supports vector-based semantic search across memory files, with configurable backends including:

| Backend | Description |
|---------|-------------|
| Built-in SQLite | Optional vector acceleration |
| QMD sidecar | Local-first search combining BM25 + vectors + reranking |
| Hybrid search | Merges both keyword and semantic signals |

### Tools

- `memory_search` - Semantic queries across memory files
- `memory_get` - Direct file retrieval