【从0到1构建一个ClaudeAgent】内存管理-上下文压缩
<p>对话一长，Token 烧得肉疼。那怎么办，做压缩</p><h2 id="java实现代码">Java实现代码</h2>
<pre><code class="language-java">public class ContextCompactSystem {
// --- 配置 ---
private static final Path WORKDIR = Paths.get(System.getProperty("user.dir"));
private static final Path TRANSCRIPT_DIR = WORKDIR.resolve(".transcripts");// 新增：对话存档目录
private static final Gson gson = new GsonBuilder().setPrettyPrinting().create();
// 压缩参数
private static final int THRESHOLD_TOKENS = 50000;// 触发自动压缩的 token 阈值
private static final int KEEP_RECENT = 3; // 保留的最近工具结果数量
// --- 工具枚举 ---
public enum ToolType {
BASH("bash", "Run a shell command."),
READ_FILE("read_file", "Read file contents."),
WRITE_FILE("write_file", "Write content to file."),
EDIT_FILE("edit_file", "Replace exact text in file."),
COMPACT("compact", "Trigger manual conversation compression.");// 新增：手动压缩工具
public final String name;
public final String description;
ToolType(String name, String description) { this.name = name; this.description = description; }
}
// ... 省略相同的 ToolExecutor 接口和基础工具实现
// --- 三层次压缩系统 ---
/**
* Layer 1: 微观压缩 - 静默替换旧的工具结果
*/
private static List<Map<String, Object>> microCompact(List<Map<String, Object>> messages) {
// 收集所有的 tool_result 条目
List<ToolResultInfo> toolResults = new ArrayList<>();
for (int msgIdx = 0; msgIdx < messages.size(); msgIdx++) {
Map<String, Object> msg = messages.get(msgIdx);
if ("user".equals(msg.get("role"))) {
Object content = msg.get("content");
if (content instanceof List) {
@SuppressWarnings("unchecked")
List<Map<String, Object>> contentList = (List<Map<String, Object>>) content;
for (int partIdx = 0; partIdx < contentList.size(); partIdx++) {
Map<String, Object> part = contentList.get(partIdx);
if ("tool_result".equals(part.get("type"))) {
toolResults.add(new ToolResultInfo(msgIdx, partIdx, part));
}
}
}
}
}
if (toolResults.size() <= KEEP_RECENT) {
return messages;
}
// 从先前的 assistant 消息中映射 tool_use_id 到 tool_name
Map<String, String> toolNameMap = new HashMap<>();
for (Map<String, Object> msg : messages) {
if ("assistant".equals(msg.get("role"))) {
Object content = msg.get("content");
if (content instanceof List) {
@SuppressWarnings("unchecked")
List<Map<String, Object>> contentList = (List<Map<String, Object>>) content;
for (Map<String, Object> block : contentList) {
if ("tool_use".equals(block.get("type"))) {
String toolId = (String) block.get("id");
String toolName = (String) block.get("name");
toolNameMap.put(toolId, toolName);
}
}
}
}
}
// 清除旧的结果（保留最近的 KEEP_RECENT 个）
List<ToolResultInfo> toClear = toolResults.subList(0, toolResults.size() - KEEP_RECENT);
for (ToolResultInfo info : toClear) {
Map<String, Object> result = info.result;
Object content = result.get("content");
if (content instanceof String && ((String) content).length() > 100) {
String toolId = (String) result.get("tool_use_id");
String toolName = toolNameMap.getOrDefault(toolId, "unknown");
result.put("content", "");// 静默替换
}
}
return messages;
}
/**
* Layer 2: 自动压缩 - 保存完整对话并生成摘要
*/
private static List<Map<String, Object>> autoCompact(List<Map<String, Object>> messages) throws IOException {
// 保存完整对话到磁盘
Files.createDirectories(TRANSCRIPT_DIR);
Path transcriptPath = TRANSCRIPT_DIR.resolve("transcript_" + System.currentTimeMillis() + ".jsonl");
try (BufferedWriter writer = Files.newBufferedWriter(transcriptPath)) {
for (Map<String, Object> msg : messages) {
writer.write(gson.toJson(msg));
writer.newLine();
}
}
System.out.println("");
// 调用 LLM 生成摘要
String conversationText = gson.toJson(messages);
if (conversationText.length() > 80000) {
conversationText = conversationText.substring(0, 80000);
}
String summary = simulateLLMSummary(conversationText);
// 用摘要替换整个对话历史
List<Map<String, Object>> compressedMessages = new ArrayList<>();
compressedMessages.add(Map.of(
"role", "user",
"content", "\n\n" + summary
));
compressedMessages.add(Map.of(
"role", "assistant",
"content", "Understood. I have the context from the summary. Continuing."
));
return compressedMessages;
}
/**
* Layer 3: 手动压缩工具
* 当 Agent 主动调用 compact 工具时触发
*/
private static String handleCompactTool(Map<String, Object> args) {
String focus = (String) args.get("focus");
String focusMsg = focus != null ? " Focus: " + focus : "";
return "Manual compression requested." + focusMsg;
}
/**
* 估算 token 数量
* 简单实现：约 4 个字符对应 1 个 token
*/
private static int estimateTokens(List<Map<String, Object>> messages) {
String messagesStr = gson.toJson(messages);
return messagesStr.length() / 4;
}
// --- 工具处理器映射 ---
private static final Map<String, ToolExecutor> TOOL_HANDLERS = new HashMap<>();
static {
// ... 省略基础工具注册
TOOL_HANDLERS.put(ToolType.COMPACT.name, ContextCompactSystem::handleCompactTool);
}
// --- Agent 主循环（集成了三层压缩）---
public static void agentLoop(List<Map<String, Object>> messages) {
while (true) {
try {
// Layer 1: 每次调用前进行微观压缩
messages = microCompact(messages);
// Layer 2: 如果 token 数超过阈值，自动压缩
if (estimateTokens(messages) > THRESHOLD_TOKENS) {
System.out.println("");
messages = autoCompact(messages);
}
// ... 省略相同的 LLM 调用逻辑
boolean manualCompact = false;
for (Map<String, Object> block : content) {
if ("tool_use".equals(block.get("type"))) {
String toolName = (String) block.get("name");
// 检查是否是 compact 工具
if (ToolType.COMPACT.name.equals(toolName)) {
manualCompact = true;// 标记手动压缩
}
// ... 执行工具
}
}
// Layer 3: 如果调用了 compact 工具，执行手动压缩
if (manualCompact) {
System.out.println("");
messages = autoCompact(messages);
}
} catch (Exception e) {
System.err.println("Error in agent loop: " + e.getMessage());
e.printStackTrace();
return;
}
}
}
// --- 辅助类和方法 ---
private static class ToolResultInfo {
int msgIndex;
int partIndex;
Map<String, Object> result;
ToolResultInfo(int msgIndex, int partIndex, Map<String, Object> result) {
this.msgIndex = msgIndex;
this.partIndex = partIndex;
this.result = result;
}
}
}
</code></pre>
<h2 id="三层次压缩系统架构">三层次压缩系统架构</h2>
<p>解决长期对话中的<strong>上下文长度限制问题</strong>，通过三层渐进式压缩策略，在<strong>不丢失关键信息</strong>的前提下<strong>大幅度缩减上下文长度</strong>，实现无限长对话的能力。</p>
<pre><code class="language-java">// 压缩流程
while (true) {
// Layer 1: 每次调用前进行微观压缩
messages = microCompact(messages);
// Layer 2: 如果 token 数超过阈值，自动压缩
if (estimateTokens(messages) > THRESHOLD_TOKENS) {
messages = autoCompact(messages);
}
// Layer 3: 如果调用了 compact 工具，执行手动压缩
if (manualCompact) {
messages = autoCompact(messages);
}
}
</code></pre>
<ul>
<li><strong>分层压缩</strong>：微观、自动、手动三层策略，粒度从细到粗</li>
<li><strong>智能触发</strong>：基于token估算自动判断压缩时机</li>
<li><strong>渐进保留</strong>：保留最近的关键信息，确保连续性</li>
<li><strong>可恢复性</strong>：压缩前保存完整对话，避免信息丢失</li>
</ul>
<h2 id="微观压缩无感地进行轻量级压缩">微观压缩：无感地进行轻量级压缩</h2>
<pre><code class="language-java">private static List<Map<String, Object>> microCompact(List<Map<String, Object>> messages) {
// 收集所有的 tool_result
List<ToolResultInfo> toolResults = new ArrayList<>();
// 保留最近的 KEEP_RECENT 个完整结果
if (toolResults.size() <= KEEP_RECENT) {
return messages;
}
// 将旧的结果替换为占位符
for (ToolResultInfo info : toClear) {
result.put("content", "");
}
}
</code></pre>
<ul>
<li>静默执行，每次 LLM 调用前运行</li>
<li>将旧的、详细的工具输出替换为简短占位符</li>
<li>保留最近的结果完整，以维持短期记忆</li>
</ul>
<h2 id="自动压缩防止上下文爆炸">自动压缩：防止上下文爆炸</h2>
<pre><code class="language-java">/**
* Layer 2: 自动压缩 - 保存完整对话并生成摘要
*/
private static List<Map<String, Object>> autoCompact(List<Map<String, Object>> messages) throws IOException {
// 1. 保存完整对话到磁盘
Files.createDirectories(TRANSCRIPT_DIR);
Path transcriptPath = TRANSCRIPT_DIR.resolve("transcript_" + System.currentTimeMillis() + ".jsonl");
try (BufferedWriter writer = Files.newBufferedWriter(transcriptPath)) {
for (Map<String, Object> msg : messages) {
writer.write(gson.toJson(msg));
writer.newLine();
}
}
// 存档保护：完整对话保存到文件，随时可查
// JSONL格式：每行一个消息，便于处理和加载
// 2. 调用 LLM 生成摘要
String conversationText = gson.toJson(messages);
if (conversationText.length() > 80000) {
conversationText = conversationText.substring(0, 80000);
}
String summary = simulateLLMSummary(conversationText);
// 3. 用摘要替换整个对话历史
List<Map<String, Object>> compressedMessages = new ArrayList<>();
compressedMessages.add(Map.of(
"role", "user",
"content", "\n\n" + summary
));
// 上下文重置：用单条消息包含存档位置和摘要
// 完整可追溯：存档路径包含在上下文中
compressedMessages.add(Map.of(
"role", "assistant",
"content", "Understood. I have the context from the summary. Continuing."
));
// 连续性保持：添加assistant确认，维持对话结构
return compressedMessages;
}
</code></pre>
<ul>
<li><strong>存档优先</strong>：压缩前先完整保存，避免信息丢失</li>
<li><strong>智能摘要</strong>：用LLM生成高质量的对话摘要</li>
<li><strong>上下文重置</strong>：大幅缩减上下文，但保留核心信息</li>
<li><strong>路径嵌入</strong>：在消息中包含存档路径，便于调试</li>
<li><strong>结构完整</strong>：保持user-assistant对话结构</li>
</ul>
<h2 id="手动压缩给予-agent-主动控制权">手动压缩：给予 Agent 主动控制权</h2>
<pre><code class="language-java">/**
* Layer 3: 手动压缩工具
* 当 Agent 主动调用 compact 工具时触发
*/
private static String handleCompactTool(Map<String, Object> args) {
String focus = (String) args.get("focus");
String focusMsg = focus != null ? " Focus: " + focus : "";
return "Manual compression requested." + focusMsg;
// Agent控制：Agent可以根据需要主动压缩
// 参数化：可以指定摘要焦点，指导LLM关注特定方面
}
</code></pre>
<pre><code class="language-java">// 在主循环中检测手动压缩调用
boolean manualCompact = false;
for (Map<String, Object> block : content) {
if ("tool_use".equals(block.get("type"))) {
String toolName = (String) block.get("name");
// 检查是否是 compact 工具
if (ToolType.COMPACT.name.equals(toolName)) {
manualCompact = true;// 标记手动压缩
}
}
}
// Layer 3: 如果调用了 compact 工具，执行手动压缩
if (manualCompact) {
System.out.println("");
messages = autoCompact(messages);
}
</code></pre>
<ul>
<li><strong>Agent自主控制</strong>：Agent可以主动管理上下文长度</li>
<li><strong>任务驱动压缩</strong>：在合适的时间点（如任务切换时）触发压缩</li>
<li><strong>聚焦摘要</strong>：可以指定摘要重点，优化信息保留</li>
<li><strong>无缝集成</strong>：与自动压缩共享底层机制</li>
</ul>
<h2 id="架构演进与价值">架构演进与价值</h2>
<p><strong>从 TaskSystem 到 ContextCompactSystem 的升级</strong>：</p>
<table>
<thead>
<tr>
<th>维度</th>
<th>TaskSystem</th>
<th>ContextCompactSystem</th>
</tr>
</thead>
<tbody>
<tr>
<td>对话长度</td>
<td>受上下文限制</td>
<td>支持无限长对话</td>
</tr>
<tr>
<td>信息保留</td>
<td>全量存储</td>
<td>智能摘要+存档</td>
</tr>
<tr>
<td>控制方式</td>
<td>被动限制</td>
<td>主动+自动压缩</td>
</tr>
<tr>
<td>长期记忆</td>
<td>任务文件</td>
<td>对话存档+摘要</td>
</tr>
<tr>
<td>上下文优化</td>
<td>无</td>
<td>三层智能压缩</td>
</tr>
</tbody>
</table>
</div>
<div id="MySignature" role="contentinfo">
<p>本文来自在线网站：seven的菜鸟成长之路，作者：seven，转载请注明原文链接：www.seven97.top</p><br><br>
来源：https://www.cnblogs.com/sevencoding/p/19821529
頁:
[1]