Agent Self-Improving Memory

Problem

Agents often confuse one-off instructions, user preferences, repeated corrections, and stale rules. Without a controlled memory path, useful lessons may be lost while irrelevant details consume context, and project-specific conventions can leak into unrelated work. This skill keeps reusable knowledge in local files, without credentials, external services, or network access.

How It Works

It captures explicit corrections, stated preferences, and repeated instructions, then prompts self-reflection after significant work. Memory is stored under ~/self-improving/ with a tiered layout: memory.md holds HOT rules, projects/ and domains/ hold scoped patterns, and archive/ holds inactive material. New lessons are not promoted immediately; they require repetition or confirmation. When a file exceeds its size limit, similar entries are merged, verbose ones are summarized, or unused patterns are demoted instead of deleted. Queries, stats, exports, and forget requests are explicit, and memory usage cites the source file.

Boundaries

It does not access calendar, email, or contacts, does not make network requests, does not read files outside ~/self-improving/, and does not infer preferences from silence or observation. It fits agents that need durable collaboration styles, project conventions, and correction history. It is not intended for external system synchronization, permission management, or automatic archival of sensitive private data.