memory

const memory: object

Type Declaration

compaction

compaction: (config?) => Middleware = memoryCompaction

Context window management with compaction strategies.

Creates a memory.compaction() middleware that manages the context window.

Automatically compacts messages before each LLM call when the token count exceeds the configured limit. Only modifies ModelContext.messages — SessionContext.history is never touched.

Five strategies from gentlest to most aggressive:

• clear-tool-results: Replace old tool results with placeholder
• truncate (default): Drop oldest messages
• window: Keep last N messages
• summarize: LLM summarizes old messages
• hybrid: Summarize old + keep recent verbatim

Parameters

config?

CompactionConfig

Working memory configuration

Returns

Middleware

Middleware that manages context window

Example

// Simple truncation (default, zero cost)
agent.use(memory.compaction({ maxTokens: 8192 }))

// Hybrid (best quality, one LLM call)
agent.use(memory.compaction({
  maxTokens: 8192,
  strategy: "hybrid",
  summaryModel: "anthropic/claude-haiku-4-5",
  keepRecentMessages: 10,
}))