Openclaw Memory Ollama

OpenClaw 本地化记忆管理系统 — 构建稳定、高效、零成本的 AI 长期记忆解决方案。基于 Ollama（本地 Embedding）+ cognitive-brain（结构化存储）+ Memory Files（日常记忆）的三档存储架构。触发场景：(1) 用户需要为 AI 助手添加记忆功能 (2) 想要本地化...

Security Scan

The skill's files describe a coherent local-memory integration (Ollama + cognitive-brain + Postgres/Redis), but the runtime instructions require editing other skills' config and local credentials while the registry metadata declares no required env/paths — this mismatch and cross-skill modification are concerning.

Purpose & Capability

The described purpose (local memory using Ollama + cognitive-brain + Postgres/Redis) is coherent with the instructions: the skill expects Ollama, a pgvector-enabled PostgreSQL, Redis, and the cognitive-brain skill. However, the registry metadata declares no required env vars, binaries, or config paths while the SKILL.md explicitly references several environment variables and specific config file locations (~/.openclaw and cognitive-brain workspace). The missing metadata is an inconsistency (documentation vs declared requirements).

Instruction Scope

SKILL.md and references instruct the user/agent to edit cognitive-brain's config.json and scripts/embed.py in ~/.openclaw/workspace-*/skills/cognitive-brain, restart OpenClaw gateway, and run verification commands. These instructions require reading and modifying files belonging to another skill and the OpenClaw installation, and expect DB credentials and local endpoints. While these actions are understandable for integration, they expand the agent's scope beyond a self-contained skill and could alter other skills' behavior.

Install Mechanism

This is an instruction-only skill with no install spec and no code executed by the registry. That minimizes supply-chain risk from downloads or packaged binaries. The deployment instructions do recommend installing Ollama via brew (official source) and pulling models; nothing is downloaded from untrusted URLs in the provided docs.

Credentials

Although the registry lists no required environment variables, the docs instruct setting OLLAMA_API_KEY and OLLAMA_HOST and show cognitive-brain config fields that require a Postgres user/password and Redis connection. Requesting DB credentials and access to other local services is proportionate to running a local memory system, but the fact these credentials and config paths are not declared in the skill metadata is a mismatch and increases risk if users supply credentials without realizing the scope. The embedding script targets only localhost (Ollama) in the docs, not external endpoints.

Persistence & Privilege

always:false and default autonomous invocation are fine. However, the instructions direct modifying other skills' configuration and the global OpenClaw config (~/.openclaw/openclaw.json), which amounts to cross-skill and system-wide changes. That level of access should be treated cautiously because it can alter agent behavior persistently.

What to consider before installing

What to consider before using/installing: - The skill is instruction-only and appears to do what it claims (local memory using Ollama + cognitive-brain + Postgres/Redis), but the registry metadata does NOT declare the environment variables or config paths the docs require — review the docs carefully before supplying secrets. - The runtime steps tell you to edit files under ~/.openclaw and the cognitive-brain skill directory (config.json, scripts/embed.py) and to modify OpenClaw's openclaw.json. Back up those files before making changes and test in a non-production environment. - The cognitive-brain config will require Postgres DB credentials and access to Redis. Prefer using a dedicated local or test database, not a production DB. - The embedding script in the docs calls http://localhost:11434 (Ollama) — there are no obvious external network endpoints, but verify scripts for any unexpected outbound connections before running. - If you run the agent with autonomy, be aware the instructions imply cross-skill file edits; consider restricting the agent's file-system privileges or performing the edits manually rather than allowing automated changes. - If anything is unclear, ask the skill author for a clear list of required env vars, exact file paths, and an explicit checklist of changes so you can audit them beforehand.

OpenClaw Memory System

本地化 AI 记忆管理系统，为 OpenClaw AI 助手提供稳定、高效、零成本的长期记忆能力。

核心架构

三层存储 + 四层记忆模型：

层级	组件	说明
接入层	飞书 / WebChat / Gateway	多渠道接入
服务层	cognitive-brain + Ollama	Memory Service + Embedding
存储层	PostgreSQL + Redis + 文件	向量 + 缓存 + 文件

四层记忆：

• L1 感官记忆 — Redis，TTL 30秒
• L2 工作记忆 — Redis，TTL 60分钟
• L3 情景记忆 — PostgreSQL，永久
• L4 语义记忆 — PostgreSQL，永久

快速开始

1. 环境要求

• OpenClaw 已安装运行
• Ollama 已安装，nomic-embed-text 模型已拉取
• PostgreSQL + Redis 服务运行中
• cognitive-brain 技能已安装

2. 配置步骤

# Step 1: 配置 Ollama 环境变量
export OLLAMA_API_KEY=ollama-local
export OLLAMA_HOST=localhost:11434

# Step 2: 配置 cognitive-brain (见 references/config-reference.md)
# 修改 config.json 中的 provider 设置

# Step 3: 重启 OpenClaw Gateway
openclaw gateway restart

3. 验证

# 验证 Embedding 服务
python3 scripts/embed.py --warmup

# 验证记忆检索
memory_search 测试查询

核心脚本

• scripts/embed.py --warmup — 预热 Embedding 服务
• scripts/embed.py <文本> — 生成文本向量

🔗 引用与致谢

直接引用的技能和代码

引用来源	引用内容	说明
cognitive-brain	brain.encode() / brain.recall() 接口、episodes/concepts 表结构、Redis 缓存逻辑	核心记忆服务架构
Ollama	nomic-embed-text 模型、/api/embeddings 接口	本地 Embedding 服务
pgvector	向量存储、ivfflat 索引、vector_cosine_ops	PostgreSQL 向量扩展

架构参考

参考来源	引用内容
lark-whiteboard	系统架构图、数据流图、四层记忆模型（小明使用此技能绘制）
ppt-generator	HTML 演示稿模板、乔布斯风设计规范

工具链

工具	用途
Ollama	本地 LLM 和 Embedding 运行服务
PostgreSQL	关系数据库 + 向量存储
Redis	热缓存、L1/L2 记忆存储
ClawHub	技能发布和管理平台
OpenClaw	AI 助手框架

开发者贡献

• 小明 (xiaoming) — 使用 lark-whiteboard 技能绘制三张架构图

参考文档

• 架构图参考 — 系统架构、数据流图、四层记忆模型
• 部署实施指南 — 完整部署步骤
• 配置参考 — 配置文件详解