Agent Memory System v4

生产级 Agent 记忆系统 — 6维坐标编码 + 语义检索 + 智能压缩

Security Scan

The skill's code, CLI commands, and storage behavior are consistent with a local agent memory system; nothing obviously requests unrelated credentials or external endpoints, but I could not review all truncated files so exercise caution and review before enabling autonomous writes.

Purpose & Capability

Name/description (production Agent memory with encoding, semantic search, compression) match the included modules (encoder, store, recall, compressor, context_builder, cli, etc.). There are no unrelated required env vars or binaries listed, and the files implement the memory features the skill claims.

Instruction Scope

SKILL.md directs the agent to run the packaged CLI (python3 ~/.openclaw/workspace/agent-memory/cli.py) and describes storing a local SQLite DB and vector cache under ~/.openclaw/workspace/agent-memory/. Those actions are appropriate for a memory system, but the runtime instructions will cause the agent to read/write user data on disk (conversation history, indexes). The skill bundle actually includes many executable Python modules (not just prose), so installing/running it will execute code on the host — confirm you trust the code. Also compression features call an LLM function (llm_fn) — ensure how that function is provided and whether it sends data to external LLM endpoints.

Install Mechanism

No install spec is present (instruction-only in registry) which is low risk, but the package includes many code files that will be executed from workspace paths. No remote downloads or obscure install URLs were observed in the provided excerpts. Verify where these files are placed and that paths referenced in SKILL.md match where the platform will store skill files.

Credentials

The skill declares no required environment variables or credentials and the code shown doesn't require cloud keys. Optional dependencies (chromadb, sentence-transformers) are reasonable for semantic search. There are no unexpected credential-like env vars requested. Still review omitted files for hidden network calls or credential reads.

Persistence & Privilege

always:false (good). The skill is intended to persist conversation data locally (memory.db, chroma_db, daily_index). It also instructs the agent to automatically write memories during conversations — this is coherent with its purpose but means the agent will store potentially sensitive chat content on disk and could write/modify many local files. Confirm you want persistent local memory and consider encryption/access controls.

Scan Findings in Context

[pre-scan-injection-none] expected: No pre-scan injection signals were detected. This is consistent with a local memory system; however many files were truncated and not fully reviewed.

Assessment

What to consider before installing/activating: - This package is a full local memory system: it will read/write a SQLite DB and vector cache under ~/.openclaw/workspace/agent-memory/ (or whatever path the platform uses). Expect persistent storage of conversation contents — review whether that is acceptable for your data sensitivity. - The skill includes runnable Python modules (CLI + many libraries). Although I saw no network endpoints or unexpected credential requests in the reviewed files, 23 files were omitted from the transcript; review all included files (especially obsidian_sync, any modules that might call external APIs or open sockets) before trusting it. - Compression uses an LLM call (llm_fn). Verify how the agent provides that function and whether using it will send memory text to an external LLM provider — if so, you may be leaking private content to that provider. - Optional dependencies like chromadb / sentence-transformers may download models or start background services; audit dependency behavior and disk/network effects. - Recommended steps: run the code in a sandboxed environment first, inspect obsidian_sync and any network/HTTP calls, check where files will be created, consider enabling encryption or restricting filesystem permissions for the memory directory, and decide whether to allow autonomous writes (agent invoking the CLI during conversation). Confidence note: I reviewed many core modules included in the manifest and SKILL.md, and found them coherent with the declared purpose. Because several files were truncated/omitted in the provided bundle, my confidence is medium; a full-line-by-line review of the remaining files would raise confidence to high.

Agent Memory Skill

为 OpenClaw Agent 提供结构化记忆能力。替代 MEMORY.md 的扁平文件方案，提供语义搜索、自动分类、记忆衰减、因果链等高级功能。

快速使用

# 写入记忆
python3 ~/.openclaw/workspace/agent-memory/cli.py remember "我决定用 Chroma 做向量库"

# 检索记忆
python3 ~/.openclaw/workspace/agent-memory/cli.py recall "用户选了什么向量库"

# 组装上下文（直接拼入 system prompt）
python3 ~/.openclaw/workspace/agent-memory/cli.py context "当前对话主题"

# 查看统计
python3 ~/.openclaw/workspace/agent-memory/cli.py stats

# 执行维护（衰减分析 + 去重 + 自修复）
python3 ~/.openclaw/workspace/agent-memory/cli.py maintain

# 压缩老记忆
python3 ~/.openclaw/workspace/agent-memory/cli.py compress

# 生成图谱
python3 ~/.openclaw/workspace/agent-memory/cli.py graph --format ascii

集成方式

对话中自动使用

当用户谈论技术决策、踩坑、偏好时，自动写入记忆：

python3 ~/.openclaw/workspace/agent-memory/cli.py remember "内容" --importance high

对话前检索上下文

在回复用户之前，检索相关记忆：

python3 ~/.openclaw/workspace/agent-memory/cli.py context "用户的问题"

Cron 定期维护

设置 cron job 每天执行维护：

• 衰减分析
• 批量去重
• 自我修复（矛盾/过时检测）

数据存储

• SQLite 数据库: ~/.openclaw/workspace/agent-memory/memory.db
• 向量缓存: ~/.openclaw/workspace/agent-memory/chroma_db/
• 质量统计: ~/.openclaw/workspace/agent-memory/quality_stats.json
• 每日索引: ~/.openclaw/workspace/agent-memory/daily_index/

依赖

Python 3.10+（已安装）。可选：chromadb + sentence-transformers（语义搜索）。