CompactionConfig

Defined in: agent/compaction/config.ts:70

Configuration for the context compaction system.

Example

// Custom configuration
const agent = await LLMist.createAgent()
  .withModel('sonnet')
  .withCompaction({
    triggerThresholdPercent: 70,
    targetPercent: 40,
    preserveRecentTurns: 10,
  })
  .ask('...');

// Disable compaction
const agent = await LLMist.createAgent()
  .withModel('sonnet')
  .withoutCompaction()
  .ask('...');

Properties

enabled?

optional enabled: boolean

Defined in: agent/compaction/config.ts:75

Enable or disable compaction.

Default

true

---

onCompaction()?

optional onCompaction: (event) => void

Defined in: agent/compaction/config.ts:127

Callback invoked when compaction occurs. Useful for logging or analytics.

Parameters

event

CompactionEvent

Returns

void

---

preserveRecentTurns?

optional preserveRecentTurns: number

Defined in: agent/compaction/config.ts:108

Number of recent turns to preserve during compaction. A “turn” is a user message + assistant response pair. Recent turns are kept verbatim while older ones are summarized/dropped.

Default

5

---

strategy?

optional strategy: CompactionStrategy | "sliding-window" | "summarization" | "hybrid"

Defined in: agent/compaction/config.ts:85

The compaction strategy to use.

• ‘sliding-window’: Fast, drops oldest turns (no LLM call)
• ‘summarization’: LLM-based compression of old messages
• ‘hybrid’: Summarizes old messages + keeps recent turns (recommended)
• Or provide a custom CompactionStrategy instance

Default

'hybrid'

---

summarizationModel?

optional summarizationModel: string

Defined in: agent/compaction/config.ts:115

Model to use for summarization. If not specified, uses the agent’s model.

Default

undefined (uses agent's model)

---

summarizationPrompt?

optional summarizationPrompt: string

Defined in: agent/compaction/config.ts:121

Custom system prompt for summarization. If not specified, uses a default prompt optimized for context preservation.

---

targetPercent?

optional targetPercent: number

Defined in: agent/compaction/config.ts:100

Target context usage percentage after compaction. The compaction will aim to reduce tokens to this percentage.

Default

50

---

triggerThresholdPercent?

optional triggerThresholdPercent: number

Defined in: agent/compaction/config.ts:93

Context usage percentage that triggers compaction. When token count exceeds this percentage of the context window, compaction is performed before the next LLM call.

Default

80