Self-Iteration Core

The problem it addresses

When an agent misses replies, repeats itself, loses context, or skips SOP steps, saying "noted" is not enough. This skill turns a failure into a checkable mechanism problem and pushes the fix to a specific rule, memory, tool, or process point.

How it works

The core loop is a four-step retrospective:
- Trace the root cause: keep asking why until the issue moves from symptom to mechanism.
- Identify the fix point: decide whether MEMORY.md, USER.md, TOOLS.md, the SOP, or a skill file should change.
- Apply the change: use edit_file for immediate-layer files, avoid overwriting with write_file, and append a change record.
- Verify and consolidate: re-run the scenario or simulate it, then promote the lesson into long-term memory.

It also defines a three-level memory lookup order: first immediate context files, then recent_memory/index.json, and finally semantic retrieval with memory_search. It fits agents that need continuously refined behavior rules. It is not meant to escalate transient tool failures, network outages, or minor user preferences into full retrospectives. Keep changes single-point so the effect of each update stays attributable.