类class io.agentscope.harness.agent.memory.compaction 2.x

CompactionConfig

public class CompactionConfig

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:61–607

压缩配置，保存相关配置项。位于 io.agentscope.harness.agent.memory.compaction 包。核心成员包括 getTriggerMessages、getTriggerTokens、getReserved、getKeepMessages、getKeepTokens、getKeepTokensMin。

Configuration for conversation compaction (summarization).

trigger — when to run compaction (by token count or message count)
keep — how many recent messages to preserve verbatim after compaction

Defaults (dynamic mode):

Trigger dynamically when estimated token count reaches model.contextWindow - reserved(20k). Falls back to {@value #FALLBACK_TRIGGER_TOKENS} tokens when the model does not report its context window, or at 50 messages (whichever comes first)
Keep tail: dynamically computed as min(8k, max(2k, usable * 0.25)) tokens. Falls back to 20 messages when the model does not report its context window
Prune: enabled by default — aggregates old tool result outputs and trims them when the prunable total exceeds 20k tokens (protects the most recent 40k tokens)
Summarization is enabled; memory flush and offload are both enabled before summary

Memory prompt landscape

The harness has three LLM-driven memory operations, each with its own prompt; they live in two complementary config classes:

Compaction summary — distills the conversation prefix into one summary message before reasoning. Prompt: #getSummaryPrompt() on this class.
Flush — extracts long-term memories into today's daily ledger. Prompt: MemoryConfig.flushPrompt().
Consolidation — periodically merges daily ledgers into MEMORY.md. Prompt: MemoryConfig.consolidationPrompt().

Configure the latter two via HarnessAgent.builder().memory(MemoryConfig...).

嵌套类摘要Nested Class Summary

类Class	类型Kind	说明Description
`Builder`	类class	构建器，用于构建并配置目标实例。 A class in io.agentscope.harness.agent.memory.compaction that persists and restores session and memory data across calls.
`TruncateArgsConfig`	类class	Truncate 参数配置，保存相关配置项。 Configuration for the lightweight argument-truncation pass that runs before summarization.
`PruneConfig`	类class	修剪配置，保存相关配置项。 Configuration for aggregate tool-result pruning.

字段摘要Field Summary

字段Field	类型Type	说明Description
`FALLBACK_TRIGGER_TOKENS` staticfinal	int	回退 TRIGGER 令牌数字段。 Fallback trigger threshold (in tokens) when the model does not report its context window and `triggerTokens` is set to dynamic mode (0).
`DEFAULT_SUMMARY_PROMPT` staticfinal	String	默认摘要提示词字段。 Default summary prompt with structured format.

方法摘要Method Summary

方法Method	说明Description
`public int getTriggerMessages()`	获取 `getTriggerMessages`。 Message count above which compaction is triggered (0 = disabled).
`public int getTriggerTokens()`	获取 `getTriggerTokens`。 Estimated token count above which compaction is triggered.
`public int getReserved()`	获取 `getReserved`。 Token buffer reserved for the compaction process itself (summary prompt + output).
`public int getKeepMessages()`	获取 `getKeepMessages`。 Number of recent conversation messages (non-SYSTEM) to preserve verbatim.
`public int getKeepTokens()`	获取 `getKeepTokens`。 Token budget for the preserved tail.
`public int getKeepTokensMin()`	获取 `getKeepTokensMin`。 Returns the keep tokens min.
`public int getKeepTokensMax()`	获取 `getKeepTokensMax`。 Returns the keep tokens max.
`public double getKeepTokensRatio()`	获取 `getKeepTokensRatio`。 Returns the keep tokens ratio.
`public String getSummaryPrompt()`	获取摘要提示词。 Prompt template used for the summarization LLM call.
`public boolean isFlushBeforeCompact()`	判断是否 `isFlushBeforeCompact`。 Whether to flush long-term memories from the prefix before compaction.
`public boolean isOffloadBeforeCompact()`	判断是否到卸载原始消息到会话 JSONL 之前压缩。 Whether to offload raw messages to the session JSONL before compaction.
`public` `TruncateArgsConfig getTruncateArgsConfig()`	获取 `getTruncateArgsConfig`。 Configuration for the lightweight pre-summarization argument truncation pass.
`public` `PruneConfig getPruneConfig()`	获取修剪配置。 Configuration for aggregate tool-result pruning.
`public` `Model getModel()`	获取模型。 Optional model override for compaction (summarization).
`public` `CompactionConfig withEffective(int effectiveTriggerTokens, int effectiveKeepTokens)`	执行 `withEffective` 方法。 Creates a resolved copy with effective trigger and keep values computed from a model's context window.
`public static Builder builder()` static	返回构建器实例。 Returns the builder.

字段详情Field Detail

FALLBACK_TRIGGER_TOKENS

public static final int FALLBACK_TRIGGER_TOKENS

staticfinal

回退 TRIGGER 令牌数字段。声明为 public static final int FALLBACK_TRIGGER_TOKENS。

Fallback trigger threshold (in tokens) when the model does not report its context window and triggerTokens is set to dynamic mode (0).

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:67

DEFAULT_SUMMARY_PROMPT

public static final String DEFAULT_SUMMARY_PROMPT

staticfinal

默认摘要提示词字段。声明为 public static final String DEFAULT_SUMMARY_PROMPT。

Default summary prompt with structured format.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:70–116

方法详情Method Detail

getTriggerMessages

public int getTriggerMessages()

获取 getTriggerMessages。

Message count above which compaction is triggered (0 = disabled).

返回Returns int

返回数值结果。Returns a numeric result.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:151–153

getTriggerTokens

public int getTriggerTokens()

获取 getTriggerTokens。

Estimated token count above which compaction is triggered.

返回Returns int

返回数值结果。Returns a numeric result.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:159–161

getReserved

public int getReserved()

获取 getReserved。

Token buffer reserved for the compaction process itself (summary prompt + output).

返回Returns int

返回数值结果。Returns a numeric result.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:167–169

getKeepMessages

public int getKeepMessages()

获取 getKeepMessages。

Number of recent conversation messages (non-SYSTEM) to preserve verbatim.

返回Returns int

返回数值结果。Returns a numeric result.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:175–177

getKeepTokens

public int getKeepTokens()

获取 getKeepTokens。

Token budget for the preserved tail.

返回Returns int

返回数值结果。Returns a numeric result.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:188–190

getKeepTokensMin

public int getKeepTokensMin()

获取 getKeepTokensMin。

Returns the keep tokens min.

返回Returns int

返回数值结果。Returns a numeric result.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:192–194

getKeepTokensMax

public int getKeepTokensMax()

获取 getKeepTokensMax。

Returns the keep tokens max.

返回Returns int

返回数值结果。Returns a numeric result.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:196–198

getKeepTokensRatio

public double getKeepTokensRatio()

获取 getKeepTokensRatio。

Returns the keep tokens ratio.

返回Returns double

返回数值结果。Returns a numeric result.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:200–202

getSummaryPrompt

public String getSummaryPrompt()

获取摘要提示词。

Prompt template used for the summarization LLM call.

返回Returns String

返回字符串结果。Returns a string.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:207–209

isFlushBeforeCompact

public boolean isFlushBeforeCompact()

判断是否 isFlushBeforeCompact。

Whether to flush long-term memories from the prefix before compaction.

返回Returns boolean

返回布尔值，表示判断结果。Returns a boolean result.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:212–214

isOffloadBeforeCompact

public boolean isOffloadBeforeCompact()

判断是否到卸载原始消息到会话 JSONL 之前压缩。

Whether to offload raw messages to the session JSONL before compaction.

返回Returns boolean

返回布尔值，表示判断结果。Returns a boolean result.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:217–219

getTruncateArgsConfig

public TruncateArgsConfig getTruncateArgsConfig()

获取 getTruncateArgsConfig。

Configuration for the lightweight pre-summarization argument truncation pass.

返回Returns TruncateArgsConfig

返回TruncateArgsConfig。Returns truncate args config.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:225–227

getPruneConfig

public PruneConfig getPruneConfig()

获取修剪配置。

Configuration for aggregate tool-result pruning.

返回Returns PruneConfig

返回PruneConfig。Returns prune config.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:232–234

getModel

public Model getModel()

获取模型。

Optional model override for compaction (summarization).

返回Returns Model

返回模型。Returns model.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:240–242

withEffective

public CompactionConfig withEffective(int effectiveTriggerTokens, int effectiveKeepTokens)

执行 withEffective 方法。

Creates a resolved copy with effective trigger and keep values computed from a model's context window.

参数Parameters

名称Name	类型Type	说明Description
`effectiveTriggerTokens`	int	effective Trigger 令牌数（类型 `int`）。The effective trigger tokens.
`effectiveKeepTokens`	int	effective Keep 令牌数（类型 `int`）。The effective keep tokens.

返回Returns CompactionConfig

返回CompactionConfig。Returns compaction config.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:248–265

builder

public static Builder builder()

static

返回构建器实例。

Returns the builder.

返回Returns Builder

返回当前实例，以支持链式调用。Returns the current instance for chaining.

源码Source io/agentscope/harness/agent/memory/compaction/CompactionConfig.java:267–269

嵌套类详情Nested Class Detail

Builder

public static class Builder

构建器，用于构建并配置目标实例。位于 io.agentscope.harness.agent.memory.compaction 包。核心成员包括 triggerMessages、triggerTokens、reserved、keepMessages、keepTokens、keepTokensMin。

A class in io.agentscope.harness.agent.memory.compaction that persists and restores session and memory data across calls. Declared in io.agentscope.harness.agent.memory.compaction.

TruncateArgsConfig

public static class TruncateArgsConfig

Truncate 参数配置，保存相关配置项。位于 io.agentscope.harness.agent.memory.compaction 包。核心成员包括 getTriggerMessages、getTriggerTokens、getKeepMessages、getKeepTokens、getMaxArgLength、getTruncationText。

Configuration for the lightweight argument-truncation pass that runs before summarization.

When triggered, large string arguments of ToolUseBlocks in older messages (before the keep window) are clipped to #getMaxArgLength() characters. This is a cheap, non-LLM operation that prevents context ballooning from verbose tool invocations (e.g., write_file, edit_file).

Defaults (when enabled via Builder#truncateArgs(TruncateArgsConfig)):

Trigger at 25 messages or 40 000 tokens
Keep the 20 most recent messages untouched
Max argument length: 2 000 characters

PruneConfig

public static class PruneConfig

修剪配置，保存相关配置项。位于 io.agentscope.harness.agent.memory.compaction 包。核心成员包括 defaults、getProtectTokens、getMinimumTokens、getMaxOutputChars、getExcludedTools、builder。

Configuration for aggregate tool-result pruning.

Prune walks backward through tool-result messages, protecting the most recent #getProtectTokens() tokens of tool output. Older tool results beyond that protection window are replaced with a head+tail preview when the total prunable amount exceeds #getMinimumTokens().

This is a lightweight, non-LLM operation that runs inside ConversationCompactor#compactIfNeeded before summarization.