Auto Summarization Loop

自动摘要循环：为长对话 AI 角色实现自动上下文管理。用于： (1) 建立多级记忆架构（核心记忆/工作记忆/长期记忆） (2) 实现滑动窗口与双水位线触发策略 (3) 异步后台压缩流程设计 (4) Persona 机器人的结构化摘要输出 适用场景：需要处理长对话、降低 API 成本、避免上下文溢出的 AI 应用

Security Scan

The skill's code, README, and runtime instructions align with its stated purpose (automatic multi-tier summarization for long conversations) and do not request unexpected credentials, installs, or external endpoints.

Purpose & Capability

Name/description describe multi-level memory and summary-trigger behaviors and the included Python module (memory_manager.py) implements exactly that: token estimation, watermark triggers, async/sync summarization, prompt construction and storage layout. No unrelated capabilities (cloud access, SSH, system-level config) are requested.

Instruction Scope

SKILL.md focuses on building prompts, triggering summaries, and integrating a user-supplied summarize_fn / call_api. It does not instruct reading arbitrary system files or exfiltrating data. It references session storage paths (a session/ structure) and calls like call_api/call_model which are placeholders — the skill expects the host to implement actual model calls.

Install Mechanism

No install spec or external downloads are provided; the skill is instruction-first with two local Python source files. There are no remote URLs, package installs, or archive extractions that would present an installation risk.

Credentials

The skill declares no required environment variables, credentials, or config paths. The code likewise does not read environment secrets or request API keys; model/API integration points are left to the host and therefore do not demand credentials from the skill itself.

Persistence & Privilege

always is false and the skill does not request persistent system-wide privileges. It stores session data in a local session/ structure (documented) and does not modify other skills or global agent settings.

Assessment

This skill appears coherent and self-contained for building summarization/memory management. Before installing: (1) verify how call_api/call_model and summarize_fn are implemented by your host — ensure they do not send data to external endpoints you don't control; (2) decide where session/ data will be stored and whether it may contain sensitive PII (long-term summary may accumulate personal data); (3) replace example persona strings (e.g., impersonating real persons) with acceptable system prompts if needed; (4) if you connect a real LLM provider, only provide credentials to the host runtime (not embedded in the skill) and audit those calls; (5) review the included test script (it adjusts sys.path) before running in production. Overall the skill is internally consistent but you should confirm the host's model-calling and storage implementations before use.

Auto-Summarization Loop - 自动摘要循环

为高拟真度 AI 角色实现自动上下文管理的完整方案。

核心架构

┌─────────────────────────────────────────────────────────────┐
│                    Prompt 构建结构                           │
├─────────────────────────────────────────────────────────────┤
│  [System Prompt]  ← 核心记忆 (永不压缩)                       │
│  ─────────────────────────────────────────────────────────  │
│  【历史摘要】  ← 长期记忆 (压缩后的摘要)                        │
│  【用户档案】  ← 结构化用户信息                                │
│  ─────────────────────────────────────────────────────────  │
│  [近期消息]    ← 工作记忆 (保留最近 N 条)                      │
│  ─────────────────────────────────────────────────────────  │
│  [用户新消息]  ← 当前输入                                     │
└─────────────────────────────────────────────────────────────┘

快速使用

1. 引入核心模块

from scripts.memory_manager import (
    AutoSummarizationLoop, 
    MemoryConfig, 
    MemoryBlock,
    TriggerType
)

2. 初始化配置

config = MemoryConfig(
    max_tokens=200000,           # 最大上下文限制
    soft_limit_ratio=0.7,       # 70% 时异步压缩
    hard_limit_ratio=0.9,        # 90% 时同步拦截
    working_messages=20,        # 保留最近 20 条
    summary_max_tokens=2000,    # 摘要最大长度
    summarize_fn=your_summarize_func  # 摘要生成函数
)

3. 处理对话

manager = AutoSummarizationLoop(config)
memory = MemoryBlock(
    core_memory="你是埃隆·马斯克...",
    working_memory=[],
    long_term_summary=""
)

# 检查触发
trigger = manager.check_watermark(memory.working_memory)

# 触发压缩
if trigger != TriggerType.NORMAL:
    await manager.handle_trigger(trigger, memory)

# 构建 Prompt
prompt = manager.build_prompt(user_message, memory)

触发策略

阈值	状态	动作
< 70%	正常	继续
70-90%	软限制	后台异步压缩
> 90%	硬限制	同步拦截压缩

摘要生成函数示例

async def summarize_fn(old_text: str, old_summary: str, max_tokens: int) -> dict:
    prompt = f"""请总结以下对话...

旧摘要：{old_summary}

{old_text}

输出 JSON：{{"summary": "...", "user_facts": {{}}}}"""

    result = await call_model(prompt, model="mini")
    return json.loads(result)

提示词模板详见 references/summary_prompts.md

存储结构

session/
├── core_memory      # System Prompt
├── working_memory   # 最近消息 (JSON 数组)
├── long_term_summary # 压缩后的摘要
├── user_facts       # 用户档案
└── last_update      # 最后更新时间

与大模型集成

async def chat(session_id: str, user_message: str):
    session = load_session(session_id)
    
    # 检查是否需要压缩
    trigger = manager.check_watermark(session.working_memory)
    if trigger != TriggerType.NORMAL:
        await manager.handle_trigger(trigger, session)
    
    # 构建消息
    messages = manager.build_prompt(user_message, session)
    
    # 调用大模型
    response = await call_api(messages)
    
    # 保存到工作记忆
    session.working_memory.append({"role": "user", "content": user_message})
    session.working_memory.append({"role": "assistant", "content": response})
    
    save_session(session_id, session)
    return response

Persona 机器人特别提示

生成摘要时确保包含：

1. 用户事实档案 - 姓名、地点、职业、偏好
2. 对话事件线 - 按时间顺序的关键事件
3. 已达成的共识 - 用户确认过的结论
4. 待跟进事项 - 未完成的任务

这样大模型在读取摘要时能更精准地把握用户特征。