Claude-Mem OpenClaw Persistent Memory Plugin
Provides persistent cross-session memory for every agent on OpenClaw gateways, automatically capturing tool usage, compressing it with AI, and injecting relevant context into future sessions.
The plugin defines clear config schema and provider selection, but key operations like observation injection, auto-install, and external service dependencies lack explicit user confirmation or least-privilege enforcement. The auto-install script runs remote install scripts and clones repos, posing supply-chain and external effects risks, but no evidence of malicious code or data exfiltration. Thus trust score is 18, deducted for lack of explicit user confirmation and isolation mechanisms.
Provides multi-tier testing (unit, smoke, E2E) and detailed installation/troubleshooting docs, but static review cannot fully verify key paths in real environments; e.g., install.sh is partially shown, and tests rely on mock worker. Thus reliability is 10, deducted for incomplete verification from static read.
Clear target scenario (persistent memory for OpenClaw gateway) with detailed config fields and trigger events, but lacks explicit capability boundaries and non-fit scenarios; core depends on external AI providers potentially unreachable in mainland China. Thus adaptability is 7, deducted for limited boundary and environmental evidence.
Well-structured docs with quick install, manual setup, config reference, troubleshooting, and testing; but version inconsistencies (plugin.json 13.13.1 vs SKILL.md 1.0.0) and missing changelog/known limits. Thus convention is 12, deducted for versioning and known-limits gaps.
Plugin appears to complete core functions (record observations, inject context, stream to channels) per docs, with example outputs and verification commands, but static review cannot verify actual result correctness or direct usability. Thus effectiveness is 7, deducted for lack of verifiable output examples and comparative benefit.
Provides test suite (src/index.test.ts), E2E scripts, and verification steps in docs, which are auditable, but static review cannot independently reproduce key behaviors. Thus verifiability is 5, deducted for lack of third-party execution evidence or cross-source corroboration.
- Core functionality relies on external AI services (e.g., Gemini, Claude) that may be unreachable from mainland China; users should assess accessibility.
- The auto-install script executes remote scripts (e.g., install.cmem.ai), posing supply-chain risks; manually review or use manual installation.
- Version inconsistencies (13.13.1 vs 1.0.0) may cause confusion and update management issues.
- The skill collects and processes tool-usage data from sessions; ensure compliance with privacy regulations and user consent.
What does this skill do, and when should you use it?
This skill enables persistent memory across sessions for agents running on OpenClaw gateways. It automatically captures tool usage (Read, Write, Bash, etc.) as observations, compresses them with AI via a local worker service, and injects relevant context into future sessions via system prompt injection. It supports real-time observation feeds to messaging channels like Telegram, Discord, Slack, and more. Ideal for users who want their agents to maintain knowledge continuity across sessions without manual memory management.
This skill operates through OpenClaw lifecycle hooks: initializes sessions on agent start, injects observation timeline into system prompts before prompt building (cached 60s), records each tool usage (truncated to 1000 chars) as observations, and summarizes sessions on agent end. It communicates with a local worker service (default port 37777) that stores observations, generates summaries, and serves context injection endpoints. An optional SSE observation feed forwards new observations to messaging channels (Telegram, Discord, Slack, Signal, WhatsApp, LINE). It registers two commands: /claude_mem_status and /claude_mem_feed.
- A developer running a coding bot on OpenClaw who wants to preserve project context across sessions, so future sessions don't need to re-learn prior decisions.
- A team monitoring agent activity in a Telegram or Discord channel, wanting to see a stream of observations in real time as tools are used.
- A user managing multiple agents with different projects, using the 'project' config to isolate memories.
- A user who wants automatic tool usage recording but prefers not to rely on MEMORY.md for context management.
What are this skill's strengths and limitations?
- Fully automatic recording with no manual intervention
- Context injected via system prompt, keeping MEMORY.md agent-controlled
- Real-time observation feed to multiple messaging channels
- Includes worker viewer UI at http://localhost:37777
- Automated install script simplifies setup
- Requires an AI provider (e.g., Gemini, Claude) for AI summarization; provider selection and costs not documented
- Observation feed only supports configured messaging channels; setup requires channel IDs, which may be non-trivial
- Automated install script is from an undocumented source (install.cmem.ai), which may raise trust concerns
- Documentation doesn't mention support for custom worker deployments beyond default port and localhost
- No test suite or clear failure modes beyond troubleshooting table
How do you install this skill?
Quick install: run curl -fsSL https://install.cmem.ai/openclaw.sh | bash. The installer checks dependencies (Bun, uv), installs the plugin, configures memory slots and AI provider, and starts the worker. Options include --provider=gemini --api-key=YOUR_KEY to preselect provider, --non-interactive for unattended, --upgrade for upgrades. Manual setup: clone repo, run npm install and npm run build, ensure Bun is installed, start worker (npm run worker:start), and add plugin config to OpenClaw gateway config.
How do you use this skill?
After installation and restart, the plugin works automatically. Agent tool usage is recorded via the lifecycle events described. Run /claude_mem_status in an OpenClaw chat to verify worker health, and /claude_mem_feed to check feed status. Optionally, enable the observation feed in plugin config by setting observationFeed.enabled, channel, and to. Then have an agent do work and observe observations stream to your channel. See SKILL.md for full config fields.