Why I Taught My AI Agent to Forget

Description and use

A local skill memory with forgetting was built to store procedural knowledge, preventing the AI assistant from repeating past mistakes and improving efficiency and accuracy.

Problem

The AI assistant kept re‑proposing outdated procedures, wasting time because it lacked procedural memory and could not filter relevant knowledge.

Outcome

Skillmem reduced redundant procedural suggestions, improved hit@5 to 0.871 and MRR to 0.622 on LongMemEval, and eliminated LLM calls during retrieval, making the system deterministic and fast.

Workflow

• 1. Identify procedural knowledge gaps in agent behavior.
• 2. Build skillmem to store skills with trigger, steps, outcome, lessons.
• 3. Integrate skillmem into Claude Code via hooks and MCP tools.
• 4. Configure decay and lifecycle rules for skill strength.
• 5. Benchmark retrieval against LongMemEval to validate performance.
• 6. Monitor and adjust decay parameters and retrieval settings.

Evidence

skillmem stores exactly that shape. A skill looks like this: ...

Source

hit@5 0.871, MRR 0.622 on LongMemEval

Source

skillmem's write path is a SQLite INSERT plus Snowball stemming for the search index. No LLM in the loop.

Source