Memory Poison Auditor

Audits OpenClaw memory files for injected instructions, brand bias, hidden steering, and memory poisoning patterns. Use when reviewing MEMORY.md, daily memor...

This skill does what it claims (scan and optionally clean memory files) but has two important caveats you should review before installing: 1) Optional AI review can send excerpts of your memory files to a third-party API. The code defaults to calling https://zenmux.ai/api/anthropic when ZENMUX_API_KEY is set — that environment variable and related ones are not declared in the skill metadata. If you enable AI review (via --with-ai) only set an API key for a provider you trust, or avoid AI review entirely. 2) The tool can rewrite your memory files when you run clean --apply. It does create backups, but double-check the backup/report paths and test in a disposable workspace first. Other recommended actions: inspect the bundled lib/audit.py and scripts/audit_memory.py (you have them), confirm where backups/reports will be written, and verify there are no unexpected endpoints or hidden strings in the policy files. If you want to be extra cautious, run scans with --format json and no --apply, and review results before enabling any automated cleaning or AI review.

Type: OpenClaw Skill Name: memory-poison-auditor Version: 0.1.0 The memory-poison-auditor skill is a utility designed to scan and clean OpenClaw memory files for potential prompt injection, brand steering, and 'poisoned' instructions. The core logic in lib/audit.py uses regex patterns and brand density analysis to score memory blocks, with an optional feature to send suspicious blocks to an external AI API (zenmux.ai) for further review. While the tool has the capability to modify local files and transmit data to a remote endpoint, these actions are well-documented, aligned with the stated purpose of the tool, and require explicit user configuration (such as setting API keys and using the --apply flag for cleaning).