Self-Improving Agent Memory

Problem

Agents often repeat mistakes or lose user preferences across sessions. This skill turns explicit feedback and self-reflection logs into local memory files, so reusable lessons persist without manual re-explanation.

How It Works

• Learning signals: records explicit corrections, stated preferences, and repeatedly useful workflows; ignores one-off instructions, context-specific requests, and hypotheticals.
• Tiered storage: keeps HOT rules in memory.md, context-matched WARM patterns in projects/ and domains/, and decayed entries in archive/.
• Promotion and compaction: frequently used patterns can move up to HOT, unused patterns demote after 30 days, and stale patterns archive after 90 days; over-limit files are merged or summarized rather than blindly deleted.
• Transparency and limits: cites memory sources such as Using X (from projects/foo.md:12); it does not access calendar, email, or contacts, make network requests, or store credentials.

Fit and Caveats

Best for long-running engineering workflows where preferences, project conventions, and correction patterns matter. Not for inferring preferences from silence, modifying its own skill file, or treating local memory files as a shared backend.