dream-memory

完整的工作区记忆管理系统。四层架构：文件存储 + OpenViking 向量引擎 + Ollama bge-m3 + Agent 规则。 Use when: 用户询问记忆系统如何工作、如何安装记忆系统到新 Agent、记忆文件结构、向量检索原理、 Session Flush 机制、长期记忆晋升规则、OpenVik...

Security Scan

The skill's files, instructions, and local checks are consistent with a workspace memory-management helper; nothing requested is disproportionate, but the reference install step (curl | sh) and local config reads deserve cautious review before running.

Purpose & Capability

The skill describes a memory stack (files + vector DB + local LLM + agent rules) and its instructions and scripts only touch local workspace files, an OpenViking vector service, and a local Ollama model—these map to the stated purpose.

Instruction Scope

SKILL.md and the self-check script explicitly read workspace files (MEMORY.md, AGENTS.md, etc.) and OpenClaw session/config files (~/.openclaw/...), which are relevant for session flush/check behavior. No instructions appear to collect or transmit data to external endpoints beyond local services.

Install Mechanism

There is no formal install spec, but references/ollama-setup.md recommends running `curl -fsSL https://ollama.com/install.sh | sh` and using `ollama pull bge-m3`. Download-and-execute via curl|sh is common for installers but is higher-risk than package installs; the URL is the official domain (ollama.com), which mitigates but does not eliminate risk.

Credentials

The skill declares no required env vars or secrets. The script optionally reads OPENCLAW_CONFIG or defaults to ~/.openclaw/openclaw.json—this is proportional to verifying memorySearch settings and session tracking.

Persistence & Privilege

The skill is not always-enabled and does not request special platform privileges. It can be invoked autonomously (normal default); combined with local-file access this is expected for a memory-management skill but users should be aware it may read agent session/config files when run.

Assessment

This skill appears coherent with its stated purpose (managing workspace memories using a local vector DB and local LLM). Before installing or running: (1) review the Ollama install script at https://ollama.com/install.sh yourself instead of piping blindly to sh; (2) be aware the skill and its self-check read ~/.openclaw/openclaw.json and agent sessions files—inspect those files for sensitive info you don't want read; (3) running ollama pull bge-m3 may download large model data and start a local service; (4) if you enable autonomous invocation, expect the skill to read and update local memory files/sessions as described. If you want higher assurance, request a signed/reviewed installer or run steps manually under an isolated account or container.

Dream Memory — Agent 记忆管理系统

架构总览四层

第 1 层: Agent 规则 (AGENTS.md) — 何时写、写哪、怎么写
第 2 层: 文件存储 (Markdown)   — MEMORY.md + memory/ + WEEKLY-PROGRESS.md
第 3 层: 向量检索 (OpenViking)  — memory/ 目录的向量化索引
第 4 层: 会话生命周期           — sessions.json 追踪 + Session Start Check

文件结构（工作区根目录）

workspace/
├── AGENTS.md             # 操作规则（记忆写入逻辑在此）
├── MEMORY.md             # 长期记忆索引（< 200 行, < 25KB）
├── SOUL.md               # Agent 人格
├── USER.md               # 用户画像
├── WEEKLY-PROGRESS.md    # 本周进度表
├── memory/
│   ├── YYYY-MM-DD.md     # 每日日志（Append only）
│   ├── topics/           # 专题详情
│   ├── archive/          # 压缩归档
│   └── .data/            # OpenViking 数据
└── .consolidate-lock     # 蒸馏检查锁

实时记录规则

何时写： 用户纠正/偏好/目标/"记住"/报告完成/外部指针 格式： - HH:MM 类型：内容 规则： Append only，不修改历史行，跨日自动切换文件

Session End Flush 🔥

触发：用户说"下班/再见"、>30 min 未写、心跳 >8h 未写、跨日切换

检查清单：重要决定 / 新知识 / 错误 / 待办 / MEMORY.md 超 4KB

输出格式：## [日期] - 主题 (标签) 每条 1-2 行

Session Start Check 🔍

每次新会话启动时：

1. 读取 ~/.openclaw/agents/{agent-id}/sessions/sessions.json
2. 找上一个用户会话的 endedAt
3. 对比当天日志最后修改时间
4. 日志时间 < endedAt → ❌ 补写

补写格式：## [日期 HH:MM 补写] — 标题\n- 修复了什么\n- 补写原因

详细配置（OpenViking + bge-m3 参数）→ 见 references/ollama-setup.md 详细 Ollama 安装指南 → 见 references/ollama-setup.md 自检脚本 → scripts/memory-check.sh