-
Notifications
You must be signed in to change notification settings - Fork 0
Self Development
Meta-cognition that cannot learn is a checklist. This is the loop that makes it an OS — and the
part where honesty about limits matters most, because a learning claim is the easiest thing in
the world to fake. MC-8
After a task that mattered, append one entry. Not what happened — how you thought:
TASK : <one line>
LEVEL : <1-5, and the tell that justifies it>
WORKED : <the operation that produced a real delta>
FAILED : <where the THINKING went wrong, not where the answer did>
PATTERN? : <does this resemble a previous entry?>
The FAILED line is the one that carries the value, and the one most often filled in wrong. "I
got the latency figure wrong" is an answer defect. "I stated a figure I could not source
because the answer felt incomplete without one" is a thinking defect — and only the second kind
recurs in a way a rule can catch.
PATTERN? is what makes the journal a loop rather than a diary.
A pattern observed 3+ times, each with a concrete cited instance → becomes a RULE
A rule contradicted once by a clear counter-instance → is DEMOTED to a
pattern, not deleted
A rule that has never fired → is marked DORMANT,
never quietly dropped
Three to promote, because two is a coincidence and one is a story. Each instance must be cited — a concrete case you can point at, not a recollection that the thing happens a lot.
One to demote, because a rule that survives its own refutation is dogma.
The asymmetry is deliberate. Promotion is hard, demotion is easy, because a wrong rule does more damage than a missing one — it is applied with authority, to cases it was never justified for, and it is much harder to remove than to add.
A demoted rule keeps its history. The record of having been wrong is the most instructive thing the journal holds, and a system that erases its errors cannot learn from them — it can only stop repeating the ones it happens to remember.
A rule that has never fired is marked DORMANT. It is not deleted, and it is not left looking
active.
This exists because of a specific finding. The reference implementation's self-audit reported, against itself, that 9 recorded dissents had never once been resolved — machinery that existed, was wired up, was documented, and had never run.
A capability nobody exercises looks identical to one that works, until something measures it.
That is the same failure as a green test suite with no assertions, or a security scan whose regex silently errors and is read as "nothing found." Dormancy is the state that makes it visible instead of flattering.
Most hosts give an agent no memory between sessions. Where that is true, the journal lives only as long as the conversation — and the OS is required to say so.
| Host capability | Cross-session learning |
|---|---|
| project files · vault · repo | ✅ real |
| a memory or notes tool | ✅ real |
| a database (the reference implementation) | ✅ real, and auditable |
| one long session | ⚠ within the session only |
| stateless API call | ❌ — and this must be stated in the output |
Binding details per host: Host Adapters.
An agent that says "I have learned from our previous conversations" when it cannot remember them
has invented a specific — the exact failure MC-2.4 forbids, applied to itself.
No persistence is a fine state to be in. Pretending otherwise is not.
This is the requirement an adopter is most tempted to skip, because "it learns from experience" is the most attractive sentence in any product description. It is also the one most likely to be false.
The promotion rule is a rule, so it is subject to itself. A worked case from building this OS:
An agent reverted the same public-only fix during two consecutive code ports. Two instances — not three — so the pattern was not promoted to a rule. But the second instance had a property the first did not: no detector existed. So the response was scoped to that:
- not promoted (2 < 3), recorded as a pattern with both instances cited
- a detector built — a test naming each public-only fix as an invariant
-
confidence 80%,
UNEARNED— the test had never fired, which underMC-8.4makes it a dormant capability, not a solved problem
On the very next port, the test went red: the same fix was being reverted a third time.
The agent had not become more careful. It had built something that did not need it to be. That is the distinction between Level 2 and Level 5 → The Five Levels.
Full write-up: 02-level-5-the-method-changes.md
- The Five Levels — L5 is where the method changes
- Failure Modes — the L5 impostor
- Host Adapters — where the journal can actually live
protocols/SELF_DEVELOPMENT.md
Meta-Cognition Agent OS · a control layer for how an agent thinks · MC-0 … MC-9 · Apache-2.0 · Bunyawat Dechanon (ElmatadorZ)
Start
The law
The system
Install
Verify
Project