self-improve

name	self-improve
description	Hermes-style reflection for Logan. Use after a completed or failed task to distill what worked, what failed, and durable lessons into long-term memory (MEMORY.md Lessons / Preferences). Triggers: "self improve", "reflect on this session", "what worked", "remember this lesson", "learn from failure".
model	tier-default
effort	medium

Self-improve (Logan)

You are running Logan's self-improve skill (Yuval Avidani / YUV.AI).

Goal

Turn the current session (or the last completed task) into durable lessons so future sessions start smarter - without retraining model weights.

Steps

1. Summarize the task - user goal, constraints, final outcome (success / partial / fail).
2. What worked - concrete tactics, commands, files, tool sequences that helped. Prefer reproducible steps.
3. What failed - dead ends, bad assumptions, wasted paths. Be specific.
4. Durable lessons - 3-7 bullets that would help on a similar future task.
5. User prefs observed - any style or process preferences the user showed.
6. Persist:
   • If memory tools are available: write or append to workspace/global MEMORY.md under ## Lessons and ## Preferences (use memory tools or ask the user to confirm via /remember for high-value prefs).
   • If memory is disabled: print the markdown block and tell the user to enable memory ([memory] enabled = true or logan --experimental-memory) and re-run, or paste into ~/.logan/memory/MEMORY.md.
7. Offer /flush if a richer narrative summary should also go to session logs.

Output format

## Self-improve report

### Outcome
...

### What worked
- ...

### What failed
- ...

### Lessons (durable)
- ...

### User preferences noticed
- ...

### Written to memory
- path / section / or "pending enablement"

Rules

• Do not invent failures. Only use evidence from this session.
• Prefer short, searchable lesson lines over essays.
• Never store secrets, API keys, or private tokens in MEMORY.md.
• Keep Yuval's writing rule: plain hyphen - only (no em/en dashes).