Description

Agent 睡眠系统 - 记忆整合、日志归档、工作区清理（支持 CortexGraph）

README (SKILL.md)

Agent Sleep System 🛌

Name: Agent Sleep
Author: guohongbin-git

像人类一样，Agent 需要"睡眠"（离线维护）来防止记忆碎片化和上下文污染。

功能

Micro-Rest - 快速上下文修剪
Deep Sleep - 每日日志整合到长期记忆
CortexGraph 同步 - 同步到 CortexGraph（带遗忘曲线）
Dreaming - 后台模拟（可选）

工具

sleep_status

检查 agent 是否"累了"（基于运行时间或 token 使用）

python3 scripts/sleep_status.py

run_sleep_cycle

触发睡眠周期

Light: 压缩最近日志
Deep: 归档 + 文件清理
CortexGraph: 同步到 CortexGraph

python3 scripts/run_sleep_cycle.py --mode [light|deep|cortexgraph]

schedule

设置生物钟（cron jobs）

python3 scripts/schedule.py --set "0 3 * * *"  # 3 AM 睡眠

工作流程

Deep Sleep 模式

1. 触发 → Cron 在 3:00 AM 启动
2. 读取 → memory/YYYY-MM-DD.md（昨天日志）
3. 提取 → 高价值洞察
4. 追加 → 到 MEMORY.md
5. 归档 → 原始日志到 memory/archive/
6. 清理 → 删除临时文件

CortexGraph 模式

1. 读取 → MEMORY.md + daily logs
2. 同步 → 到 CortexGraph
3. 应用 → 遗忘曲线（自动衰减）
4. 晋升 → 高价值记忆到 LTM

遗忘曲线

CortexGraph 使用 Ebbinghaus 遗忘曲线：

score = (use_count)^β × e^(-λ × Δt) × strength

β = 0.6（使用频率权重）
λ = ln(2) / half_life（默认 3 天）
strength = 1.0-2.0（重要性）

使用

手动触发

# 轻量睡眠
python3 scripts/run_sleep_cycle.py --mode light

# 深度睡眠
python3 scripts/run_sleep_cycle.py --mode deep

# CortexGraph 同步
python3 scripts/run_sleep_cycle.py --mode cortexgraph

定时设置

# 每天凌晨 3 点深度睡眠
python3 scripts/schedule.py --set "0 3 * * *"

# 每 6 小时 CortexGraph 同步
python3 scripts/schedule.py --set "0 */6 * * *"

目录结构

agent-sleep/
├── SKILL.md
├── AGENT.md
├── scripts/
│   ├── run_sleep_cycle.py
│   ├── sleep_status.py
│   └── schedule.py
└── memory/
    ├── archive/        # 归档的日志
    └── consolidated/   # 整合的记忆

配置

环境变量

# CortexGraph 配置
export CORTEXGRAPH_STORAGE_PATH=~/.config/cortexgraph/jsonl
export CORTEXGRAPH_DECAY_MODEL=power_law
export CORTEXGRAPH_PL_HALFLIFE_DAYS=3.0

ClawHub 配置

clawhub install agent-sleep

最佳实践

每日 Deep Sleep - 凌晨 3 点
每 6 小时 CortexGraph 同步 - 保持记忆新鲜
每周 GC - 清理低分记忆
每月晋升 - 高价值记忆升级到 LTM

与其他 Skill 集成

Skill	集成方式
memory-sync-cn	使用其脚本同步到 CortexGraph
agent-library	使用其压缩功能
cortexgraph	直接调用 MCP 工具

版本: 1.1.0 更新: 添加 CortexGraph 支持

Usage Guidance

Do not install or schedule this skill for automatic runs until the author clarifies and fixes inconsistencies. Specific concerns to resolve before use: 1) Missing/incorrect files referenced in SKILL.md/README (sleep_status.py, schedule.py, src/*) — ask for a corrected, minimal feature set. 2) The run_sleep_cycle.py provided only implements a non-destructive 'light' routine; confirm and review implementations for 'deep' and 'cortexgraph' modes before trusting archival or deletion behavior. 3) Remove or make configurable the hard-coded path in scripts/note.py (/Users/guohongbin/...) — it will write to another user's directory if present; require a configurable NOTES_PATH. 4) Ensure any file-moving or deletion operations are explicit, reversible (use archive/ and safe backups), and operate only inside the agent's workspace. 5) If you plan to enable CortexGraph sync, require and validate explicit CORTEXGRAPH credentials/paths (don't rely on implied environment variables). 6) Run this skill in a sandbox or inspect/execute the scripts manually with a test workspace before granting it access to your real agent data or scheduling it.

Capability Analysis

Type: OpenClaw Skill Name: agent-sleep Version: 1.1.0 The skill is classified as suspicious primarily due to the presence of `scripts/note.py`, an unannounced file unrelated to the 'agent-sleep' functionality. This script contains a hardcoded absolute path (`/Users/guohongbin/mcp-note-taker`) for storing notes, which is a significant vulnerability as it makes the script non-portable and could lead to unintended file system interactions if executed in an environment where this specific path exists. Additionally, there are notable functional discrepancies between the skill's documentation (SKILL.md, AGENT.md, README.md) and the actual implementation in `scripts/run_sleep_cycle.py`, which only supports a 'light' sleep mode and lacks the described deep sleep, archiving, cleanup, and CortexGraph synchronization features.

Capability Assessment

⚠ Purpose & Capability

The README/SKILL.md claim Deep Sleep, archival, CortexGraph sync, scheduling, and integration with other skills, but the included code only implements a trivial light sleep reader (scripts/run_sleep_cycle.py) and a 'note' helper. CortexGraph behaviour, archive/move operations, schedule/sleep_status scripts, and the promised semantic compression are absent. Several references point to non-existent files/paths (e.g., sleep_status.py, schedule.py, src/run_sleep_cycle.py). This mismatch suggests the manifest/documentation are out-of-sync with the implementation.

⚠ Instruction Scope

SKILL.md and AGENT.md instruct reading/writing memory/YYYY-MM-DD.md, appending to MEMORY.md, moving logs to memory/archive/, deleting temp files, and scheduling cron jobs. The actual run_sleep_cycle.py only reads a daily file from ~/.openclaw/workspace/memory and prints stats — it does not perform archival, append to MEMORY.md, apply forgetting curves, or call CortexGraph. The instructions therefore ask the agent to manipulate files and schedules that the code does not implement; following the docs blindly could result in missed operations or unintended file access if modified.

✓ Install Mechanism

This is an instruction-only skill with no install spec; nothing is downloaded or installed automatically. That limits immediate supply-chain risk. However, shipped scripts will run if the user invokes them locally.

⚠ Credentials

No required env vars are declared, yet SKILL.md suggests CORTEXGRAPH_* environment variables for CortexGraph integration — those are optional in docs but not used in code. More importantly, scripts/note.py contains a hard-coded absolute path to /Users/guohongbin/mcp-note-taker/notes.txt and will append to that file if present. That is disproportionate and fragile: it assumes another user's path and will write to an unrelated location without configuration, which is unexpected for this skill.

✓ Persistence & Privilege

The skill is not always:true and does not request elevated platform privileges. It does not modify other skills' config in the provided code. The main risk is local file I/O performed by scripts when run.

Version History

v1.1.0

**CortexGraph support and workflow enhancements** - Added support for CortexGraph memory synchronization, including forgetting curve models. - Expanded sleep modes: now includes "light", "deep", and "cortexgraph" options. - SKILL.md fully rewritten in Chinese, with detailed workflow, configuration, and best practices. - Updated tools/scripts references and directory structure for easier integration. - Introduced configuration options for CortexGraph decay model and scheduling. - Added package.json for skill packaging.

v1.0.0

Initial release of agent-sleep – a biological-inspired rest & memory consolidation system for Agents. - Enables micro-rest (context pruning), deep sleep (memory consolidation and log archiving), and optional dreaming (background simulation) - Provides tools to check tiredness, manually trigger sleep cycles, and schedule automatic sleep with cron - Automates log consolidation into long-term memory and cleans up excess files - Helps prevent memory fragmentation and context pollution for agent systems

Metadata

Slug agent-sleep

Version 1.1.0

License —

All-time Installs 3

Active Installs 3

Total Versions 2

Frequently Asked Questions

What is Agent Sleep?

Agent 睡眠系统 - 记忆整合、日志归档、工作区清理（支持 CortexGraph）. It is an AI Agent Skill for Claude Code / OpenClaw, with 700 downloads so far.

How do I install Agent Sleep?

Run "/install agent-sleep" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Agent Sleep free?

Yes, Agent Sleep is completely free (open-source). You can download, install and use it at no cost.

Which platforms does Agent Sleep support?

Agent Sleep is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Agent Sleep?

It is built and maintained by Guohongbin (@guohongbin-git); the current version is v1.1.0.

More Skills

Agent Sleep