Agent Self-Evolving Core

What it solves

A common failure mode for agents is not weak reasoning but missing durable experience: similar mistakes recur, constraints stay buried in chat, and session knowledge disappears after restart. self-evolving-core places task execution, experience extraction, rule prechecks, and memory maintenance into one closed loop, so known pitfalls become searchable, reusable, and verifiable memory.

How it works

• Task logging: after each task, log records the work and automatically checks rules R1-R4, triggering review, immune-rule extraction, or distillation prompts.
• Automatic review: review extracts success patterns and failure lessons from task traces, writes them to experiences.json, and preserves fuller scenario and lesson text.
• Layered memory: memory is split into declarative, procedural, semantic, and working layers; MEMORY.md should hold rules, indexes, and pointers rather than raw experience text.
• Immune prechecks: the Router checks existing rules before task execution and emits critical, warning, or info guidance to reduce repeat mistakes.
• Dual-location sync: core scripts and config must stay consistent between the working copy and skill source to avoid new workspaces inheriting old bugs.

Boundaries

It fits agent workflows with clear domains and long-term task experience. Benefits are limited for one-off tasks, weakly abstractable lessons, or unstable directory structures. Adopting it means treating log as a mandatory accounting step and maintaining the PINNED section and dual-location sync rules in MEMORY.md.