KIP

KIP 2.0 — Cognitive State Protocol for Agent Memory

[English](./README.md) | [中文](./README_CN.md)

Status

Default development version: KIP 2.0 — normative draft. The repository root contains the current protocol and tooling. KIP v1.0-RC11 is frozen for historical reference and migration; changing the default does not upgrade existing deployments or declare the 2.0 protocol stable.

• The scope is frozen. A new contract enters the draft only together with engine evidence; corrections, simplifications and new evidence do not need that gate (Specification Status).
• Names are stable; revisions are digests. There is one memory package, kip://profiles/cognitive-memory@2.0.0, one general domain package, kip://domains/general@1.0.0, one standard projection policy, kip:memory-default, one standard strength policy, kip:strength-half-life-30d, and schema IDs urn:kip:2.0:schema:*. A draft revision is identified by its content digest; earlier draft packages are not retained.
• Both reference engines pass the whole engine suite — 431 cases in 26 fixtures, none pending. The suite manifest records the verified engine revision.
• No learning result is claimed. Engine runs, bounded models and structural tests are separate evidence from measured Brain behavior, which stays not_run until a Brain evaluation measures it.

What the latest revision changed, and how it was validated, is in the revision record.

Start here

• Agent integration: Memory Interface and the Agent memory card.
• Protocol implementation: Specification, syntax reference, and conformance checks.
• Running a Brain: Anda Brain, a memory service that serves the Memory Interface; see Implementations.
• Developer tools: TypeScript language toolkit (@ldclabs/kip-lang) and VS Code extension, both targeting KIP 2.0.
• Existing v1 users: frozen archive, deployment migration, and repository path changes.

The archived MCP server and Skill implement v1 interfaces; native v2 integration starts with the current Interface and wire contracts above.

An interactive version — pan, zoom, relationship tracing, five guided views — is a self-contained file: download it and open it in a browser.

The one idea

A memory that cannot separate *what was said* from *what is true* will eventually lie to its owner with total confidence. KIP 2.0 refuses to collapse three questions:

Meaning      what can be represented at all

Belief what the Brain currently accepts Authority who may read, write, project, or elevate

So a stored Proposition is truth-neutral. An Assertion carries one actor's stance toward it, with mode, confidence, validity and evidence. Belief is not a stored field but an Epistemic Projection computed at read time over the eligible assertions:

Proposition exists  ≠  Proposition is true  ≠  Brain accepts Proposition

Change never rewrites the past. Three kinds of change keep three different histories:

• The world changed. One new Assertion from the time of the change. Temporal succession closes the actor's older value at that point, so a formerly true claim is ended rather than declared wrong, and historical queries still answer with it. Two inferences never succeed one another.
• A claim was wrong. A new Assertion supersedes the old one, which stays on record.
• The Brain misrecorded it. Recording repair invalidates the wrong extraction; the source bytes and everyone's stance stay as they were.

Time is stated as honestly as it is known: asserted_at is when a claim was made, not when it was recorded; a claim with no stated start began no later than that; and a coarse date is a time bound, not an invented instant. A preference is an ordinary claim that changes within its kind — there is no Preference type. Where claims of one slot still conflict, the standard memory policy kip:memory-default resolves them the same way on every engine: context specificity first, then first-person testimony, then recency.

Evidence, provenance and schema get the same treatment: Evidence and Activity are first-class elements rather than metadata bags, and schema lives in versioned, digest-pinned packages. A Brain that meets a relation the Profile lacks drafts it into the Space's draft vocabulary with DEFINE, which only adds; promoting a draft onto an installed symbol is a Schema migration that only a Principal holding manage_schema performs. Retrieval composes with belief in one query through the Search Pattern: search finds candidates, and BELIEF decides what is accepted.

The proactivity layer

A memory that only answers when asked is half a memory. The other half is what the Brain does when nobody is talking to it — and it must not be a cron job. KIP 2.0 gives that half four mechanisms, none of which decides policy on the Brain's behalf:

Supporting state lives in the Cognitive Memory Profile: WorkingState is the consolidated resume digest stamped with its basis_seq, so a Brain wakes from compiled state plus a delta rather than re-reading scrollback; computed dependency validity makes a revised root visible at the next read; and MnemonicState.utility holds the admission bet — how useful this memory is expected to be — kept deliberately separate from salience, from memory_strength, and from epistemic confidence. Strength itself is computed at read time from a base, an anchor and a pinned decay policy; nothing sweeps it, and reading never reinforces it.

The consequence channel

Everything above makes the system watch the world better. The consequence channel is how the world watches back.

Outcome Evidence records what actually happened after a decision, action, or trialed procedure — written by instrumentation (telemetry, verifiers, test harnesses, human review), never by the actor whose action it grades. An actor's own account is agent_statement, citable as context only; the separation is a conformance invariant, enforced as auditability — engine origin always records who wrote what, and Governance can restrict who may write outcomes — because an open protocol can make self-grading visible even where it cannot make it impossible.

Each outcome carries a task family that selects candidate consequences. Sharing the family establishes neither attribution nor baseline membership: under the optional Validated Learning companion, TrialRecord explicitly freezes comparable baseline attempts and outcomes. Treatment observations link through the instrument's outcome_observation Activity to an attempt and the action_gate decision that applied the exact Skill revision. Its DecisionRecord distinguishes retrieval from actual use, and grading counts independent attempts assigned before execution. A SkillRevision must name its task family before trial; a pattern that nothing could prove wrong is not procedural memory. On the channel sits the Skill lifecycle:

proposed → trialed → adopted → revoked

Lifecycle changes commit with a validated immutable EvaluationRecord on a lifecycle_verdict Activity and a guarded update (Validated Learning §7). Auditors replay exact inputs against the immutable TrialRecord; a Skill's current_trial only selects it, and its GradingState is a computed view. Without the companion, Skills stay unproven candidates — useful, recallable, never promoted. Only trialed → adopted promotes through a comparative verdict, and revocation is never harder than adoption. Same-state monitoring may retain standing under authorized policy without claiming new improvement; policy withdrawal may have zero outcom