Does it automatically remove duplicate information?

Yes, observations with over 90% similarity to recent memories are rejected, and the existing memory's helpful count and recency are updated instead of creating a duplicate.

How does the memory feedback loop work?

The skill allows you to provide helpful or harmful signals to memories. A helpful signal increases importance, while harmful signals decrease it, effectively suppressing low-quality context without immediate deletion.

What is effective importance?

It is a calculated score ranging from 1 to 10 that combines a memory's base importance with user feedback to determine how likely it is to be retrieved.

Can I prevent memories from being deleted?

You can disable the pruning process by setting the TTL (Time-To-Live) to zero, or ensure memories stay by marking them with high importance scores.

Engram Memory Curation

Name: Engram Memory Curation
Author: grootenberg

bygrootenberg

0•

Productivity & Workflow

Manages AI memory quality through feedback loops, automated deduplication, and importance calibration to ensure relevant context.

The Engram Memory Curation skill enhances Claude's performance by actively managing its internal knowledge system to prevent context noise and information decay. It implements a sophisticated feedback mechanism that allows the AI to learn from its successes and mistakes, automatically adjusting the 'effective importance' of stored memories based on their utility. By utilizing cosine-similarity for deduplication and heuristic-based importance defaults, the skill ensures that only high-quality, relevant information is retained while pruning stale or misleading data. This is an essential tool for developers who want to maintain a clean, efficient, and highly accurate long-term memory for their AI assistant.

Key Features

01Automated cosine-similarity deduplication to prevent redundant context

020 GitHub stars

03Heuristic importance defaults for instructions, errors, and code changes

04Detailed health monitoring to track memory distribution and feedback ratios

05Feedback-driven importance calibration using helpful and harmful signals

06Configurable memory compaction and TTL-based pruning for automated maintenance

Use Cases

01Streamlining AI responses by deduplicating redundant insights across long-term projects

02Optimizing retrieval accuracy by penalizing outdated or misleading AI memories

03Maintaining a high-signal context by automatically pruning low-importance episodic data

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add grootenberg/engram memory-curation

For use in Claude.ai and ChatGPT

Download Skill