memory-referee

Memory hygiene and adjudication layer for OpenClaw agent workflows. Deduplicates entities, resolves naming conflicts, separates facts from goals from specula...

This skill appears coherent and self-contained: it adjudicates an input array of memory records in-process and returns a report and structured JSON. Before installing, consider: 1) Review the included source (you already have it) or run the tests to verify behavior matches your expectations — classification, similarity threshold (0.8), and staleness TTL (30 days) are heuristic and may need tuning. 2) It does not persist data or contact external services by default, so secrets are not requested or used. 3) The repository contains developer governance notes (CLAUDE.md referencing a 'saturnday' tool) — these are developer workflow instructions and do not execute during normal runtime, but follow them only if you intend to modify the code. 4) For very large record sets (>100k), the README warns this is in-process only and not optimized for streaming; consider batching or an external pipeline for scale. If you want higher assurance, run the test suite (npm test), inspect the compiled dist files that will be executed, and optionally run a static scan for any changes before use.

Type: OpenClaw Skill Name: memory-referee Version: 0.1.0 The memory-referee skill is a utility for processing and cleaning agent memory records through deduplication, keyword-based classification (fact vs. goal vs. speculation), and heuristic conflict detection. The code is well-structured, includes a comprehensive test suite, and operates entirely in-memory without any network or sensitive file system access. No evidence of malicious intent, data exfiltration, or prompt injection was found; the 'Saturnday' branding and governance files (CLAUDE.md, saturnday-state.json) appear to be part of the AI-assisted development workflow used to generate the skill.