Where are the evaluation records stored?

All evaluations are stored locally on your machine as JSONL files within the ~/.claude/alm/evaluations/ directory.

How does overriding an evaluation affect Claude?

When you override an evaluation, the system recalculates the confidence score for that task type, which directly impacts Claude's autonomy level for future similar tasks.

How do I view a specific session's evaluation?

You can pass a specific session ID as an argument to the review command to retrieve and display the detailed evaluation record for that specific interaction.

Can I fix a 'false positive' correction detection?

Yes, the skill provides an override function to set correctionDetected to false if the ALM misclassified a standard request as a user correction.

What is the primary purpose of the Review skill?

It allows you to audit and correct the data stored in the Active Learning Memory (ALM) system to ensure Claude learns accurately from your feedback and interactions.

ALM Learning Review

Name: ALM Learning Review
Author: sfw

bysfw

0•

分析与监控

Manages and corrects Claude Code's learning evaluations to improve system accuracy and autonomy scores.

The ALM Review skill serves as the primary feedback mechanism for the Claude Code Learning Memory System, allowing users to audit and refine how the model learns from interactions. It provides a structured way to view recent task evaluations, identify misclassified corrections, and override outcome statuses. By providing this human-in-the-loop validation, users can directly influence the confidence scores and autonomy levels of the AI, ensuring that the system's internal memory remains accurate and that Claude evolves correctly based on real-world usage.

主要功能

01Manually update task classifications and success outcomes

020 GitHub stars

03Override false positive or false negative correction detections

04View detailed logs of recent task evaluations and session IDs

05Recalculate task-specific confidence and autonomy scores in real-time

06Directly edit local JSONL evaluation records via terminal automation

使用场景

01Adjusting task types for better categorization of future automated workflows

02Correcting a misidentified 'failure' outcome to prevent unnecessary confidence score drops

03Auditing recent interactions to ensure the ALM system is capturing learning moments correctly

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add sfw/alm review

For use in Claude.ai and ChatGPT