Fluid Memory

基于艾宾浩斯遗忘曲线和访问频率的衰减模型设计的遗忘和归档机制，完全依赖openclaw原生记忆系统的拟人化流体记忆系统

Security Scan

The skill's files and runtime instructions are consistent with a local, file-based memory/archive tool using ChromaDB; nothing in the bundle looks purposely malicious, but there are privacy and robustness concerns you should be aware of before installing.

Purpose & Capability

Name/description (a local 'fluid' memory system) matches what is included: Python scripts, a hook that logs messages, and optional ChromaDB integration. Declared dependencies in SKILL.md (python, chromadb, pyyaml) align with the code. No unrelated cloud credentials or unrelated binaries are requested.

Instruction Scope

SKILL.md instructs the agent to call local CLI wrappers (python wrapper.py ...) and to rely on OpenClaw's memory flush. The included hook writes every user message to a local conversation_log.txt — this is within a memory skill's scope but is an important privacy surface: messages are logged in plaintext and persist locally until processed/archived. The SKILL.md states auto_learn=false, but the hook still records all messages to disk awaiting flush — this is functionally consistent but worth highlighting.

Install Mechanism

There is no remote install/download in the spec — the skill is instruction-and-code-only. All code is bundled in the package; no external URLs or extract steps are present. Risk from install mechanism is low.

Credentials

The skill requests only a Python runtime (declared) and the SKILL.md lists python packages chromadb and pyyaml. No environment variables, keys, or unrelated credentials are requested. This matches the stated purpose.

Persistence & Privilege

The skill does persistent local storage under ~/.openclaw/workspace/database/ and includes a daemon (dream_daemon.py / maintenance.py) that will archive and hard-delete items based on thresholds. always:false and user-invocable:true are appropriate. The hook (if enabled) will log every 'message:sent' event — so the skill can collect/retain conversational content passively on the host; this is expected for a memory skill but increases privacy exposure.

Assessment

This skill appears to do what it says: a local, ChromaDB-backed memory system that logs conversations and applies decay/archive rules. Key things to consider before installing: 1) Privacy: every 'message:sent' may be appended to ~/.openclaw/workspace/database/conversation_log.txt in plaintext; sensitive data will be stored locally unless you disable the hook or clear files. 2) Encryption/permissions: if others share the machine, enable disk encryption and restrict file permissions on the workspace directory. 3) Chroma dependency: without chromadb installed the skill degrades or returns errors for major features — ensure you install the declared Python packages or accept limited functionality. 4) Hook behavior: the included hook will actively log messages when enabled; disable the hook if you don't want passive logging. 5) Review/verify code: one of the source files showed a truncation marker in the provided package listing — inspect the full fluid_skill.py before trusting behavior. 6) Robustness/platform notes: the code uses Windows-style paths in places and has fallbacks for different Chroma APIs; test on your environment. If any of these concerns are unacceptable, do not install or disable the hook and/or delete the local data directory after uninstall.

Fluid Memory Skill

这是你的「赛博大脑」。它不是死板的数据库，而是一个活着的系统——会遗忘不重要的事，会强化常被提及的知识。

自动学习模式 (Auto Learn)

通过 OpenClaw 原生 flush 触发！ 每次 OpenClaw 触发 memory flush 时，AI 会同步调用 fluid-memory 记录对话。

• 依赖 OpenClaw 原生 compaction 机制（配置 softThresholdTokens 控制频率）
• 需在 OpenClaw 配置中启用 memoryFlush

遗忘机制

• 动态遗忘：检索时分数 < 0.05 被过滤
• 主动遗忘：调用 fluid_forget 归档指定记忆
• 梦境守护：定时归档分数 < 0.15 的记忆

核心理念

• 植入 (Remember): 写入新记忆。
• 唤起 (Recall): 检索记忆。每次检索都会强化该记忆（访问次数+1）。
• 遗忘 (Forget): 将匹配的记忆归档。

工具 (Tools)

你 (OpenClaw LLM) 可以调用以下工具来与我（Fluid Memory）交互：

1. 植入记忆

当用户明确说「记住xxx」时使用。

Trigger: 用户说「记住...」「植入...」「记录...」 Tool Call:

{
  "name": "fluid_remember",
  "arguments": {
    "content": "用户说的原话"
  }
}

2. 唤起记忆

当用户问「我之前说过什么」「记得xxx吗」「我的偏好是什么」时使用。

Trigger: 用户问「我喜不喜欢...」「还记得...」「我之前...」 Tool Call:

{
  "name": "fluid_recall",
  "arguments": {
    "query": "用户的查询关键词"
  }
}

3. 遗忘

当用户说「忘记xxx」「不要再提xxx」时使用。

Tool Call:

{
  "name": "fluid_forget",
  "arguments": {
    "keyword": "要遗忘的关键词"
  }
}

4. 状态

查看大脑状态。

Tool Call:

{
  "name": "fluid_status",
  "arguments": {}
}

内部实现 (供开发者参考)

实际执行命令：

python wrapper.py remember --content "..."

隐私

• 数据存储在本地 ~/.openclaw/workspace/database/
• 明文存储（无加密）
• 无云端同步

最佳实践

1. 自然触发: 不要机械地调用工具，要理解用户的意图。
2. 强化重要记忆: 如果用户多次提到某事，调用 fluid_recall 查询它，这会让它记得更牢。
3. 接受遗忘: 如果系统返回「没有相关记忆」，不要慌，这说明用户很久没提这件事了，或者确实没说过。