Three Tier Memory

三级记忆管理系统 (Three-Tier Memory Management)。用于管理 AI 代理的短期、中期、长期记忆。包括：(1) 滑动窗口式短期记忆，(2) 自动摘要生成中期记忆，(3) 向量检索长期记忆 (RAG)。当需要管理对话历史、优化上下文、构建个人知识库、或实现记忆持久化时使用此 Skill。

Security Scan

Skill generally implements a three-tier memory manager that matches its description, but there are several inconsistencies in its instructions vs. code (config format, LLM integration), and the script reads/writes a workspace path via an undeclared WORKSPACE_DIR environment variable — review before installing.

Purpose & Capability

The code implements short/medium/long-term memory (sliding-window JSON, summaries, and a local ChromaDB vector store) which matches the skill's stated purpose. Using a local vector DB (Chroma) and local files is reasonable for this purpose.

Instruction Scope

SKILL.md and references instruct running the included Python script and mention YAML config and external LLM models, but the script: (a) actually saves config as JSON (not YAML), (b) implements a placeholder local summarize function instead of calling an LLM, and (c) the behavior writes files to a workspace directory — these mismatches mean the runtime behavior may differ from user expectations. The SKILL.md also suggests using specific models (e.g., 'glm-4-flash', 'gpt-3.5-turbo') though the script does not perform real LLM calls.

Install Mechanism

No install spec is provided (instruction-only + included script). That is low-risk in terms of install mechanism because nothing is fetched during install; code is shipped with the skill.

Credentials

The skill declares no required env vars, but the script reads WORKSPACE_DIR (defaulting to '/Users/scott/.openclaw/workspace') to determine where it writes memory files. This environment dependency is not documented in SKILL.md. No credentials or secret env vars are requested, which is proportionate.

Persistence & Privilege

The skill does not request always: true and does not modify other skills or system-wide settings. It persists data under a workspace directory (creates files and directories), which is expected behavior for a memory manager.

What to consider before installing

This skill appears to implement the advertised three-tier memory system, but there are several mismatches between the documentation and the code you should review before installing: (1) The docs ask for a YAML config but the script uses a JSON config file; (2) The docs mention automatic summarization via LLMs, yet the script currently uses a local placeholder summary routine (no LLM network calls) — if you expect integrated LLM summaries you must inspect/modify code to provide the intended API hooks and ensure credentials are handled safely; (3) The script writes files into WORKSPACE_DIR (default /Users/scott/.openclaw/workspace) but the SKILL.md does not declare or highlight this environment variable — set WORKSPACE_DIR to an isolated directory or inspect the default path before running; (4) The long-term store uses chromadb if installed; installing third-party Python packages should be done in a virtualenv and reviewed. Recommendation: review the included scripts/memory_manager.py source yourself (or run it in an isolated environment), confirm where files will be written, and only enable LLM/network integrations after verifying how credentials would be provided and stored. If you need higher assurance, request a version that actually integrates with your intended LLM backend and documents required env vars and install steps.

Memory Manager Skill

管理 AI 代理的三级记忆系统：短期（滑动窗口）、中期（自动摘要）、长期（向量检索）。

快速开始

# 初始化记忆系统
python3 scripts/memory_manager.py init

# 添加短期记忆
python3 scripts/memory_manager.py add --type short --content "用户喜欢黑色"

# 查询记忆
python3 scripts/memory_manager.py search "用户的偏好"

架构概览

层级	存储位置	触发条件	用途
短期	memory/sliding-window.json	实时	保持当前对话连贯
中期	memory/summaries/	Token 阈值	压缩历史，保留大意
长期	memory/vector-store/	语义检索	永久记忆，RAG

核心功能

1. 短期记忆：滑动窗口

• 配置：config/window_size（默认 10 条）
• 逻辑：FIFO 队列，超出则丢弃最旧消息
• 文件：memory/sliding-window.json

2. 中期记忆：自动摘要

• 触发：当前 token > config/summary_threshold（默认 4000）
• 模型：使用廉价模型（如 GPT-3.5-Haiku）
• 输出：memory/summaries/YYYY-MM-DD.json

3. 长期记忆：向量检索

• 后端：ChromaDB（本地向量库）
• 存：对话结束/摘要生成后自动向量化存储
• 取：每次查询前先检索相关记忆

配置文件

创建 memory/config.yaml：

memory:
  short_term:
    enabled: true
    window_size: 10
    max_tokens: 2000

  medium_term:
    enabled: true
    summary_threshold: 4000
    summary_model: "glm-4-flash"  # 或 gpt-3.5-turbo

  long_term:
    enabled: true
    backend: "chromadb"
    top_k: 3
    min_relevance: 0.7

使用场景

• 新对话开始：先 search 长期记忆，注入相关上下文
• 对话中：自动管理短期/中期记忆，超阈值自动摘要
• 对话结束：将重要信息存入长期记忆

详细用法

See REFERENCES.md for complete command reference.