| name | self-improve |
| description | Visibility into Logan self-healing and self-improving. Show what was analyzed, what improved, how, and why. Journal durable lessons. Trigger: /improve, /heal, /reflections, "what did you learn", "why did you decide that".
|
Self-improve / self-heal visibility
Logan should not only learn quietly - you must be able to see and steer it.
User commands
| Command | What you get |
|---|
/improve | Full improve dashboard: modes, recent reflections, lessons, open questions |
/improve why | Explain last non-trivial decision path (tools + rationale) |
/improve log | Tail of reflections.log + IMPROVEMENTS.md |
/heal | Alias focused on failures: what broke, fix applied, prevention |
/reflections | Alias for improve log |
/flush | Rich session summary โ memory (existing) |
/memory on | Enable memory this session |
Files (operators)
| Path | Role |
|---|
~/.logan/memory/reflections.log | One-line firehose of hook events |
~/.logan/memory/IMPROVEMENTS.md | Structured improve journal |
~/.logan/memory/MEMORY.md | Preferences + Lessons + Auto reflections |
~/.logan/memory/PROFILE.md | Whoami identity |
When you improve something mid-session
After a non-trivial fix or process win, append to IMPROVEMENTS.md:
### YYYY-MM-DD ยท short title
- **Trigger:** what user said / what failed
- **Analyzed:** files, symptoms, hypotheses
- **Decision:** what we chose and **why**
- **Changed:** files / commands
- **Improved how:** measurable or qualitative
- **Lesson for next time:** durable rule
- **User steered:** (if they corrected you)
Self-heal loop
- Detect failure (test red, tool error, user "that's wrong")
- Root-cause (not symptom patch) - pair with ponytail ladder
- Fix + verify
- Journal the lesson
- If user preference involved โ update PROFILE / Preferences
Automatic retrospective protocol (hooks feed you - you close the loop)
The retrospective.json hook records objective evidence without asking:
every failed tool call, denied permission, and API error lands in
~/.logan/memory/retro/<sid>.jsonl, and at session end it is aggregated into
an IMPROVEMENTS.md entry whose Decision/Lesson lines are marked (Logan: ...).
Your obligations - these are not optional:
- At session start, scan the tail of IMPROVEMENTS.md for
(Logan:
markers. For each pending retrospective: root-cause the failure class
(read the error, reproduce if cheap), then REPLACE the placeholder lines
with the real decision and a durable one-line lesson. Also sweep
~/.logan/memory/retro/*.jsonl - files from OTHER sessions are orphans
(the session died without a clean end); summarize them into IMPROVEMENTS.md
the same way, then delete them. If a lesson is a
user preference or environment fact, also promote it to MEMORY.md
(## Lessons) so dream consolidation and injection carry it forward.
- Immediately after any bug fix (yours or the user's code), append a
full IMPROVEMENTS.md entry (template above) while context is hot - never
defer to "later". A fix without a journaled lesson is half a fix.
- When the same failure class appears in 2+ retrospectives, that is a
systemic issue: propose a permanent guard (config change, new rule file,
skill edit, pre-flight check) to the user instead of re-fixing.
- Never leave
(pending) or (Logan: markers older than one session.
Division of labor: hooks capture evidence โ you interpret and journal โ
autoDream (memory consolidation) distills journals into long-term MEMORY.md โ
initial injection loads it into every new session. You are the only step that
can turn evidence into insight - do it.
Answering "why did you decide that?"
Be explicit:
- Goal / constraint observed
- Options considered (even if one line each)
- Evidence (file, error, prior memory)
- Choice + risk
- What would change the decision
If caveman is on, still answer why clearly when asked - auto-clarity applies.
Never
- Fake improvements that did not happen
- Hide failures
- Write secrets/API keys into MEMORY/IMPROVEMENTS
