memory-proposer

To append learned patterns safely, this Skill generates draft PRs that restrict edits to the "Learned Patterns" section of SKILL.md files. It isolates changes to skills or references directories, ensuring agent code remains untouched.

Pattern detection works by clustering active and promoted memory cases using embedding similarity. It identifies recurring patterns and maintains identity stability across re-runs by matching new clusters to existing ones via centroid similarity.

Yes, you can propose detected patterns as draft PRs using content-hash branch names. The Skill enforces safety guardrails by always creating draft PRs, ensuring proposed updates are reviewable before merging.

Yes, scipy and numpy are required dependencies for clustering memory cases by embeddings. These libraries provide the numerical and scientific computing foundation needed for calculating embedding similarity and centroid matching.

The best way to review detected memory patterns is using the list command. After detecting clusters with the detect command, you can review them and then generate a draft PR for a chosen pattern with the propose command.

You should not use automated draft PRs if you need immediate changes to skill files, as they are created as drafts requiring manual review. Additionally, edits are restricted only to the "Learned Patterns" section of SKILL.md files.