Summarization and context limits

Summarization changes what the next model call sees. It does not replace the complete stored transcript used by a conversation UI. Choose a strategy through history.WithSummarizer when constructing the conversation manager. Import pkg/agents/history and pkg/agents/history/summariser (the package uses the spelling summariser).

Keep a sliding window

A sliding window retains recent runs without making another model call:

window := summariser.NewSlidingWindowHistorySummarizer(
    &summariser.SlidingWindowHistorySummarizerOptions{KeepCount: 10},
)
manager := history.NewConversationManager(
    history.NewInMemoryConversationPersistence(),
    history.WithSummarizer(window),
)

The count is in runs, not individual messages. Keeping complete runs preserves tool-call/result groupings. A non-positive count is floored to one. Older context is omitted, not compressed into a summary, so the model can lose facts that are not in the retained window.

Summarize older runs with a model

Given a configured model and persistence adapter:

summarizer := summariser.NewLLMHistorySummarizer(&summariser.LLMHistorySummarizerOptions{
    LLM: model,
    Instruction: hastekit.NewPrompt("Summarize decisions, constraints, and unresolved tasks accurately."),
    TokenThreshold: 24000,
    KeepRecentCount: 5,
})
manager := history.NewConversationManager(persistence, history.WithSummarizer(summarizer))

TokenThreshold is a context-occupancy trigger, not a lifetime token budget. The manager combines the latest provider usage with estimates for messages added since that call. KeepRecentCount preserves recent runs unchanged; despite historical comments referring to messages, grouping is by run. Its default is five. Leave room below the selected model’s context limit for instructions, tools, new input, and the next response. Tool results can enlarge context between model calls. Summarization itself costs a model call and can fail.

Custom strategies and filters

Implement history.HistorySummarizer for another strategy. Its SummaryResult identifies the retained messages and summarization boundary. A history.MessageFilter can alter the messages selected for model input. Keep related tool calls and outputs together, preserve unresolved interruptions, and test continuation after a summary boundary. Use the transcript APIs for display rather than reconstructing chat history from the shortened model context.

Getting started

Core concepts

Models and providers

Tools and capabilities

Multi-agent patterns

Interfaces

Durable execution

Scheduled work

Workflows

Testing and operations

Summarization and context limits

Keep a sliding window

Summarize older runs with a model

Custom strategies and filters

​Keep a sliding window

​Summarize older runs with a model

​Custom strategies and filters

Keep a sliding window

Summarize older runs with a model

Custom strategies and filters