Memento

Local persistent memory for OpenClaw agents. Captures conversations, extracts structured facts via LLM, and auto-recalls relevant knowledge before each turn....

What to consider before installing: - Defaults are privacy-first, but extraction that sends text to cloud LLMs is opt-in (extraction.autoExtract defaults to false). Keep autoExtract off if you do not want any conversation text sent to external providers. - If you want fully air-gapped operation, run a local Ollama model and configure extractionModel to an ollama/* model; no cloud API key needed. - Migration is powerful: the migrate tooling can read workspace files specified in ~/.engram/migration-config.json or via MEMENTO_WORKSPACE_MAIN. Only run migration if you trust the configured paths and have reviewed which files will be ingested (these can include large or sensitive local files). - Data is stored at ~/.engram/conversations.sqlite and ~/.engram/segments/*.jsonl — inspect, back up, or encrypt these files if needed. - The plugin delegates model routing to OpenClaw when run inside the platform and may use platform tokens (CLAUDE_CODE_OAUTH_TOKEN) for routing; verify your OpenClaw auth policy if you want to limit which models/providers are used. - The install is via npm (@openclaw/memento). As with any third-party package, consider reviewing the package source or installing in a sandbox before granting it access to production workspaces. Confidence note: High — the repository, SKILL.md, and changelog are internally consistent. The main risks are user-configured behaviors (enabling autoExtract or running migration) rather than silent or unexpected access.

Type: OpenClaw Skill Name: memento Version: 0.6.0 The OpenClaw Memento plugin is designed for local, privacy-first memory management, but exhibits 'suspicious' characteristics due to its broad local filesystem scanning and inherent risks of sending data to external LLMs. The `src/cli/deep-consolidate.ts` script broadly scans `homedir()` for agent databases in `~/.openclaw/workspace-*/` and `~/*/` (non-dot directories), which, while intended for legitimate discovery, represents a wide scope of local file access. Additionally, while transparent and opt-in, the plugin's core functionality in `src/extraction/extractor.ts` and `src/consolidation/relation-sweep.ts` involves sending conversation segments or fact summaries to external cloud LLM providers (e.g., Anthropic, OpenAI) if configured, which inherently carries data exfiltration risks, even with explicit privacy safeguards like `secret` fact filtering in `src/extraction/classifier.ts` and `relation-sweep.ts`. These capabilities, while plausibly aligned with the stated purpose, introduce high-risk behaviors that warrant a 'suspicious' classification rather than 'benign' due to the potential for misuse or unintended data exposure.