LESSONS.md

# Semantic Lessons

This file contains distilled lessons that tend to hold across tasks.
Entries here are promoted only after recurring evidence.
This is not a log.

---

## System & Agent Architecture

- Externalized memory is more reliable than model context for long-term behavior.
- The harness should be thin; intelligence must live in files, not hidden state.
- Assumptions about structure cause drift; explicit maps prevent it.

---

## Memory Discipline

- Working memory must be treated as disposable and task-scoped.
- Episodic records are valuable only when they surface patterns.
- Promote to semantic memory only when a lesson recurs or has high impact.
- Mixing personal preferences with general lessons causes confusion.

---

## Safety & Control

- Permissions and constraints must be enforced outside skills.
- If an action is blocked, the correct response is to stop and ask.
- Systems become unsafe when agents infer rules instead of reading them.

---

## Evolution Rules

- Skills should describe destinations and constraints, not step-by-step procedures.
- Self-rewrite must be conservative and evidence-driven.
- Memory systems should decay or archive stale information, not accumulate indefinitely.