Direct KIP: Maintenance card

English | 中文

For a model implementing Brain Maintenance. Use only the capabilities and permissions actually advertised. Load the Recall/Formation cards for shared operations and the complete syntax only when an uncommon operation needs it.

• Basic memory: consolidate supported claims, review contradictions/dependencies, retain useful evidence and keep processing receipts truthful.
• Experience: reconstruct process and compile immutable unproven candidates.
• Learning: only with memory_learning/full-profile support, schedule the authorized deterministic evaluation over independent attempts and retained trial inputs.
• Durable work: only with the worker capability, claim/renew fenced leases, consume complete Watch watermarks and reconcile the same external attempt after crashes.
• Exchange/erasure: use governed plans and verify coverage; no implicit permission.

A source revision leaves history intact and changes computed dependency validity. Review all required dependents with bounded, checkpointed traversal:

LIST DEPENDENTS :revised_root DEPTH 2 LIMIT 100

Follow continuation pages/depths as required. This first page is not a complete review. Stored DerivationState is a review record, never an override of computed validity or of the action gate's current checks.

Mnemonic state can change without changing truth confidence:

UPDATE :element
SET FACET "MnemonicState" {
  memory_strength: :new_strength, last_metabolized_at: :cycle_start
}
EXPECT VERSION :mnemonic_version OF FACET "MnemonicState"

Use a bounded policy and a stable cycle identity; re-read on conflict. A time-based sweep never rewrites Assertion confidence. Utility calibration names its evidence and attribution method, not mere retrieval exposure.

Skill behavior changes create/select a new SkillRevision. Trials and evaluations are immutable; current tallies cache their exact revision/evaluation. Do not hand write adopted status. All planes changed by a verdict are guarded in one transaction. A basic deployment without learning support simply retains unproven experience.

Refresh WorkingState for its actor/task/context from a pinned basis. Preserve the full basis behind any compact memory result. An available processing receipt records a historical completed horizon; later corrections, policy changes and erasure still control what can be recalled now.

Archive preserves history. Payload purge, semantic erasure and retraction have different meanings. Explicitly report partial/blocked plans and unavailable replay inputs. Never mark a pending input processed or a partial erasure completed to make health metrics look better.