Can I manually correct a wrong memory?

Yes, you can use the 'ccmem correct ' CLI command or the Python API to decrease a memory's confidence score and mark it for review.

How does the confidence decay work in this skill?

Memories lose confidence by roughly 1% per day if not accessed. However, memories with high positive reinforcement ratios decay significantly slower than unverified ones.

Does this skill detect feedback automatically?

Yes, it includes logic to analyze observation data for phrases like 'use X instead' or 'exactly,' allowing it to update memory confidence implicitly during a chat session.

What happens when a memory's confidence drops too low?

Once a memory's confidence falls below 0.1, its status is changed to 'archived,' preventing it from cluttering the AI's active context window.

Memory Feedback Management

Name: Memory Feedback Management
Author: gaodes

bygaodes

0•

Productivity & Workflow

Refines AI memory accuracy by processing user feedback, confidence scoring, and automated reinforcement signals.

The Memory Feedback skill provides a sophisticated framework for managing and refining Claude’s persistent knowledge base within the ccmem ecosystem. It allows users to explicitly reinforce correct behaviors or correct inaccuracies, adjusting confidence scores that dictate how and when memories are used. By implementing features like confidence decay for stale data and automatic feedback detection from conversation context, this skill ensures that Claude's localized knowledge remains relevant, accurate, and aligned with current user preferences and modern workflows.

Key Features

01Explicit Positive Reinforcement: Boost confidence scores for helpful or correct memories to ensure they are prioritized.

02Memory Supersession: Seamlessly replace old instructions with updated versions while maintaining a historical audit trail.

03Negative Feedback & Correction: Decrease confidence or flag memories for review when they become outdated or incorrect.

040 GitHub stars

05Automatic Feedback Detection: Parse natural language responses to implicitly update memory reinforcement counts.

06Confidence Decay Logic: Automatically lower the priority of unused memories over time to prevent reliance on stale data.

Use Cases

01Automating the lifecycle of localized knowledge using time-based confidence decay to keep the context window clean.

02Correcting recurring AI hallucinations or mistakes by flagging specific 'under review' memories for manual adjustment.

03Maintaining up-to-date project preferences, such as switching from npm to pnpm across multiple sessions.

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add gaodes/cc-ecosystem memory-feedback

For use in Claude.ai and ChatGPT

Download Skill