Universal Self-Reflection and Learning Memory System

Problem it addresses

AI agents often mix one-off corrections with durable preferences: mistakes are not retained, repeated instructions keep appearing, and project style can conflict with global rules. memorypor stores the material that should be remembered in local files instead of relying on session context. It is triggered by user corrections, post-task review, or detected output improvements.

How it works

• memory.md holds confirmed hot-layer rules.
• corrections.md holds explicit corrections and candidate lessons.
• projects/, domains/, and archive/ hold scoped patterns and decayed items.

Learning signals include explicit corrections such as “No, that’s wrong” and preferences such as “Always do X”. After a pattern appears three times, it asks for confirmation before promotion. Capacity limits trigger merging, downgrading, or archiving rather than blind deletion. It can also support periodic heartbeat checks, statistics, export, and import.

Boundaries

It only maintains local memory. It does not access calendars, email, or contacts, does not make network requests, and does not read files outside ~/self-improving/. It does not infer preferences from silence or hypothetical discussion, and it does not modify its own SKILL.md. It suits agent workflows that need cross-session style, process corrections, and explicit preferences.