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

# Context Compression

> Token-aware compression with intelligent retention strategies

Synthra's context compression engine intelligently manages conversation context windows. Automatically triggered when approaching token limits.

## Compression Strategies

### FIFO (First In, First Out)

Removes oldest messages. Simple and fast (under 10ms).

```typescript theme={null}
agent.setContextConfig({
  retentionPolicy: 'fifo',
  maxTokens: 4096
});
```

### Priority-Based

Removes lowest priority messages. Ideal for complex conversations (under 30ms).

```typescript theme={null}
agent.setContextConfig({
  retentionPolicy: 'priority',
  maxTokens: 8192
});
```

### Semantic Retention

Removes redundant content using embeddings. Best for long-form content (under 50ms).

```typescript theme={null}
agent.setContextConfig({
  retentionPolicy: 'semantic',
  maxTokens: 16384
});
```

## Token Limits

| Tier       | Context Window | Compression Threshold |
| ---------- | -------------- | --------------------- |
| Free       | 4,096 tokens   | 80% (3,277 tokens)    |
| Pro        | 8,192 tokens   | 80% (6,554 tokens)    |
| Enterprise | 32,768 tokens  | 85% (27,853 tokens)   |

## Manual Compression

```typescript theme={null}
await agent.compressContext(session.id);

const stats = await agent.getCompressionStats(session.id);
console.log(`Removed: ${stats.messagesRemoved}`);
console.log(`Saved: ${stats.tokensSaved} tokens`);
```

## Best Practices

| Scenario                 | Strategy |
| ------------------------ | -------- |
| Customer support         | FIFO     |
| Multi-turn conversations | Priority |
| Content generation       | Semantic |

<Note>
  Compression is non-destructive when archiving is enabled. Removed messages can be retrieved from the archive.
</Note>

<Warning>
  Semantic compression adds 20-30ms latency. Use FIFO or priority for latency-sensitive applications.
</Warning>