Memory

Memory management middleware. Prevent context window overflow by compacting conversation history with 5 built-in strategies.

memory.compaction()

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

function memoryCompaction(config?: CompactionConfig): Middleware

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

// Hybrid: summarize old + keep recent verbatim (best quality)
agent.use(memory.compaction({
  maxTokens: 8192,
  strategy: "hybrid",
  summaryModel: mySummaryModel, // LanguageModelV3 instance
  keepRecentMessages: 10,
}))

Hooks: model — checks token count and compacts before next().

Five strategies (gentlest to most aggressive):

Strategy	Description	Cost
clear-tool-results	Replace old tool results with placeholders	Free
truncate (default)	Drop oldest messages	Free
window	Keep last N messages	Free
summarize	LLM summarizes old messages	1 LLM call
hybrid	Summarize old + keep recent verbatim	1 LLM call

Config options:

Option	Type	Default	Description
maxTokens	number	8192	Token limit for context window
strategy	CompactionStrategy	"truncate"	Compaction strategy
keepLast	number	20	For window: keep last N messages
keepLastToolResults	number	3	For clear-tool-results: keep N recent results
keepRecentMessages	number	10	For summarize/hybrid: keep N recent messages
summaryModel	LanguageModelV3	—	For summarize/hybrid: model for summaries
tokenCounter	TokenCounter	chars/4	Token estimation function