The normative Cognitive Consistency contract binds final belief, immutable Skill revisions, independent attempts, replayable trials/evaluations, dependency validity, identity repair and durable workers. Lifecycle counters aggregate attempts; unlinked family outcomes are never automatically controls. Stored summaries are used only with a validated computation basis.

English | 中文

Status

Reference Cognitive Architecture / Brain-Layer Design

This document defines how a Brain can use KIP 2.0 and Cognitive Memory Profile 2.0 to learn from experience. It is not a KIP Core requirement. Normative protocol semantics come from KIP-2.0-SPECIFICATION.md.

0. Central Thesis

Knowledge  = compressed reusable regularity
Event      = what happened
Experience = the path traversed while pursuing a goal
Skill      = experience compiled into reusable policy
Memory     = how past state conditions future computation
Learning   = durable context-appropriate behavior change caused by prior cognition

Knowledge is the compression of experience; Skill is the compilation of experience; Memory is the mechanism that lets experience continue to shape the future.

1. Layer Boundary

KIP 2.0                     cognitive primitives, history, Governance, transactions
Cognitive Memory Profile    Event / Experience / Skill / mnemonic structures
Experience Learning         learning loop
Formation/Recall/Maintenance concrete Brain policy

The protocol provides signals; the Brain owns cognitive policy.

2. What Learning Is Not

A write, embedding, retrieval, summary, confidence update, or Skill object creation is not by itself proof of learning.

Strong functional test:

future behavior with relevant memory
    >
future behavior after relevant-memory ablation

If deleting an item can never change relevant prediction/decision/action, it behaves as archive rather than functional memory.

3. Learning Loop

Environment / Human / Tool
          ↓
      Observation
          ↓
       Evidence
          ├────────→ Proposition → Assertion → Epistemic Projection
          ↓
     Event / Experience
          ├────────→ Semantic Consolidation → reusable Assertion / Insight
          ├────────→ Reflection → SelfModel
          └────────→ Procedural Consolidation → Skill
                                             ↓
                                        Action Recall
                                             ↓
                                      Future Decision
                                             ↓
                                      External Action
                                             ↓
                                      Outcome Evidence
                                             └────↺

External action is outside KIP rollback. Intent and outcome are recorded around the external effect as separate cognitive transactions.

4. Experience as Trajectory

E = (g, b0, a0, o1, b1, a1, o2, ..., y, δ)

g goal, b compact state/belief context, a action, o observation, y outcome, δ feedback/surprise/prediction error.

Store only useful, observable, permitted process information. Hidden chain-of-thought is unnecessary.

5. Event vs Experience

	Event	Experience
Question	What happened?	What path was traversed?
Size	compact	multi-step
Actions	optional	first-class
Observations	optional	first-class
Failure/recovery	summary	structurally important
Main use	episodic recall	transfer/learning
Consolidation	semantic	semantic + procedural

Formation should create Experience selectively.

6. Formation Threshold

Favor Experience when there is multi-step goal pursuit, meaningful failure/recovery, prediction error, strategy revision, human corrective feedback, unusual tool result, costly/high-impact outcome, reusable sequence, or a counterexample to an existing Skill.

Routine repetition with no new signal may not deserve a new Experience. Transaction retry is not repeated Experience.

7. Evidence First

Observed input should enter as Evidence before truth-sensitive belief:

user message
→ Evidence(user_statement)
→ Proposition(Alice, prefers, DarkMode)
→ Assertion(asserted_by=Alice, mode=stated)

This separates authentication, semantic attribution, proposition meaning, and later accepted belief.

8. Prediction Error

expected observation ≠ actual observation

is a strong learning signal. It can reveal invalid assumptions, missing preconditions, hidden state, environment changes, incorrect Skill applicability, or knowledge gaps.

Surprise is not truth confidence.

9. Belief Revision

old Assertion
+ new Evidence
→ new Assertion
+ optional supersession
+ belief_revision Activity

Do not rewrite the old Assertion's confidence, rewrite Proposition tuple, or delete prior Evidence. Third-party disagreement is normally contradiction, not supersession.

10. Semantic Consolidation

Question: What reusable declarative regularity should the Brain now treat as supported?

Experiences/Evidence
→ candidate Proposition
→ derived Assertion
→ semantic_consolidation Activity
→ Epistemic Projection

Derived cognition cannot manufacture independent corroboration from multiple summaries of one root.

11. Procedural Consolidation

Question: Given similar contexts, what action policy appears to work?

successful Experiences
+ failed Experiences
+ counterexamples
→ contrast
→ conditions/procedure/failure modes
→ proposed Skill (with its task_family)
→ trial

Contrast is essential to avoid over-broad procedure learning.

12. Contrastive Learning

Compare success vs failure, same goal with different initial state, same action with different observation, same condition under different tool/environment, and same Skill with different outcome. Seek discriminating conditions rather than popularity.

13. Failure Is First-Class

Failed Experience may teach negative preconditions, diagnostic branches, recovery strategies, invalid assumptions, counterexamples, unsafe actions, and tool limitations. Some failures have greater learning value than routine successes.

14. Four Learning Products

Semantic learning

Future belief changes through new Evidence/Assertions/conflict resolution.

Mnemonic learning

Future recall changes through memory_strength/salience. Truth does not change.

Procedural learning

Future action policy changes through Skill, GradingState, applicability, and counterexamples.

Self-model learning

Future decisions change because the Brain's model of its own capabilities, limitations, preferences, identity continuity, or strategies changed. SelfModel cannot grant authority.

15. Orthogonal Signals

Signal	Question
Assertion confidence	Strength of this Assertion's stance
source trust	Reliability of source in context
memory_strength	Cognitive accessibility
salience	Importance/noteworthiness
utility	Procedural usefulness
validity/currentness	Applicability in time

Never implement not recalled recently → lower confidence without new epistemic Evidence.

16. Reinforcement vs Evidence

Repeated retrieval does not create Evidence, increase confidence, or prove truth. Independent repeated observations may increase epistemic support. Repeated successful Skill use may increase procedural utility. Repeated exposure to the same root does not create independent corroboration.

17. Skill Model

A useful Skill carries applicability, preconditions, procedure, success criteria, failure modes, counterexamples, supporting Experiences, validation history, utility, and descriptive status.

18. Skill Lifecycle

Reference cognitive lifecycle:

proposed → trialed → adopted → revoked

Lifecycle changes and grading refreshes commit with immutable, runtime-validated EvaluationRecord (Spec §15.7, Profile §14). Only trialed → adopted promotes through a comparative trial; revoked re-entry first opens a new trial. Same-state monitoring follows the authorized policy and preserves prior adoption evidence without inventing new improvement. Policy withdrawal may have zero outcomes, and revocation is never harder than adoption. This is descriptive cognition. Governance influence/authority is separate: descriptive, advisory, behavioral, executable.

19. Skill Grading

After Skill use, capture context, whether preconditions held, selected procedure, outcome, feedback, and unexpected observations.

Classify success under matching conditions, failure under matching conditions, failure under non-matching conditions, and unknown outcome. Matching-condition failure is a strong negative signal and may narrow the Skill or, through a verdict, demote it to re-trial. Grading aggregates independent attempts assigned to the exact revision/trial before execution. Instrumented Outcome Evidence links through an outcome_observation Activity to the attempt and its action_gate decision; the agent's own account is agent_statement, never a grade. Sharing a task family establishes neither treatment attribution nor control membership: comparable baseline attempts/outcomes are explicitly selected and frozen in TrialRecord.

20. Action Recall

Ordinary recall asks what do I know? Action Recall asks what past cognition is relevant to choosing the next action under the current state?

Recommended briefing:

Goal
Current state
Accepted knowledge
Contested assumptions
Unverified preconditions
Applicable Skills
Skill utility/status/authority
Supporting successful Experiences
Relevant failed Experiences
Counterexamples
Open Commitments
Constraints
Warnings

Similarity is only one signal; applicability comes first.

21. Retrieval for Learning

Ranking may consider semantic relevance, goal/state similarity, precondition compatibility, tool/environment compatibility, outcome polarity, surprise, learning value, memory_strength, salience, recency/currentness, and Governance visibility. No universal scalar is required.

22. Counterexample Retrieval

When recalling a Skill, actively look for matching failures, known failure modes, contested assumptions, negative feedback, and recent invalidation. A high-similarity success is not sufficient when a high-value counterexample exists.

23. Self-Model Learning

SelfModel updates should be slower than Event formation. Candidate sources include repeated behavior, explicit correction, high-salience Experience, validated capability gain, stable communication preference, recurring limitation, and stable mission/value signal.

24. Maintenance

Maintenance performs semantic/procedural consolidation, identity resolution, mnemonic metabolism, retention review, conflict discovery, Skill review, SelfModel refresh, Commitment review, and quarantine review under Governance.

25. Forgetting

epistemic retraction/supersession
mnemonic weakening
archive
Governance exclusion
tombstone
physical purge

are different mechanisms, not one delete-old-memory action.

26. Experience Compression

Compression may reduce trace size but should preserve goal, important state transitions, actions, observations, outcome, failure/recovery, surprise, counterexamples, and provenance lineage. Summary does not create new Evidence roots.

27. Cross-Agent Learning

Remote Experience remains remote autobiography. A local Brain may derive a new local Skill from remote Experience while preserving provenance. Remote $self, trust, or Skill authority must not auto-transfer.

28. Evaluation

Evaluate semantic retention, temporal evolution, Experience reconstruction, procedural transfer, error avoidance, context discrimination, negative transfer, and causal memory impact.

The strongest causal experiment is with relevant memory versus relevant memory ablated.

29. Transaction Boundaries

Use atomic Transactions where partial state would mislead:

Evidence + Proposition + Assertion
Experience + Steps + Formation Activity
new Assertion + supersession + revision Activity
Skill + compiled_from + compilation Activity

External action remains outside KIP transaction rollback.

30. Idempotency

Use transaction idempotency_key, durable client_key, and stable source-event identity. Same retry ≠ new observation.

31. Provenance Conservation

Every transformation should allow the Brain to recover what inputs caused the derived cognition, whether they were observed/stated/inferred/imported, how many independent roots exist, and which Principal performed the transformation.

32. Authority Conservation

Useful/adopted/derived cognition cannot raise authority by itself. Imported Experience → local proposed Skill → local adoption still does not imply tool permission.

33. No Hidden Chain-of-Thought Requirement

Use observable action, observation, outcome, feedback, and concise decision summaries. Private token-level reasoning is not required for useful procedural memory.

34. Reference Brain Cycle

WAKE
  Formation → Evidence/Event/Experience/Assertions/Commitments
  Recall → Grounding/BELIEF/Experience/Skill/Action Briefing

SLEEP
  Consolidate → Insight/Assertion/Skill
  Metabolize → memory_strength/salience/retention
  Review → conflicts/identities/Skills/SelfModel/Commitments

NEXT WAKE
  changed cognitive state influences behavior

35. Design Invariants

The protocol-level invariants live in Specification §102 and the Profile-level ones in Profile §23, both registered in one list, KIP-2.0-Invariants.md; this document does not restate them. What it adds is the learning-specific residue:

1. A write is not proof of learning.
2. Prediction error is not confidence.
3. Similarity is not applicability.
4. Retrieval is not functional memory unless it can influence future cognition.
5. Learning should be behaviorally evaluable — with the relevant memory against its ablation.

36. Final Principle

A learning Brain is not one that remembers more. It is one whose past can change its future in the right contexts without falsifying where that past came from.