Openclaw Memory

Agent memory with ALMA meta-learning, LLM fact extraction, and full-text search. Observer calls remote LLM APIs (OpenAI/Anthropic/Gemini). ALMA and Indexer w...

What to consider before installing: - The Observer component will send conversation text to third-party LLM endpoints (OpenAI, Anthropic, Gemini) and therefore may expose sensitive or private conversation content. Only enable it if you trust the target LLM and are comfortable with that data flow. - You must supply an API key (OPENAI_API_KEY / ANTHROPIC_API_KEY or pass apiKey in config) for the Observer to work. The registry metadata does NOT declare these required env vars — do not assume the skill is fully offline. - If you don't want networked extraction, you can still use ALMA and the Indexer offline; disable or avoid instantiating ObserverAgent. - Prefer giving a least-privileged key (limited scope or separate account) if possible, and audit any logs at your LLM provider for unexpected requests. - The Indexer reads specific workspace Markdown paths (MEMORY.md, memory/YYYY-MM-DD.md, bank/entities/*.md, bank/opinions.md). Ensure those files don't contain secrets you don't want included in search or sent to the Observer. - The codebase is small and matches its README, but the metadata mismatch about required credentials is a real coherence issue — consider asking the publisher to update the skill manifest to declare the required env vars before installing.

Type: OpenClaw Skill Name: openclaw-memories Version: 2.1.0 The OpenClaw Memory skill bundle is classified as benign. It clearly states its functionality, including the use of remote LLM APIs for the Observer component and local file system access for the Indexer. Crucially, the `src/observer.ts` file includes a `sanitize` function specifically designed to prevent invisible prompt injection, and the `src/scanner.ts` file is a dedicated Unicode smuggling scanner, indicating a strong focus on security and prevention of prompt injection attacks. The database (`src/db.ts`) is an in-memory mock, mitigating traditional SQL injection risks. There is no evidence of data exfiltration, unauthorized persistence, arbitrary code execution, or other malicious intent. The package also lists zero runtime dependencies, reducing supply chain risk.