Tell a typical agent memory "I live in Tokyo," then "I moved to Berlin," and ask where you lived in March. It will say Berlin, with total confidence. KIP 2.0 is an open protocol for agent memory built so that failure can't even be expressed. The 7-minute video walks through it. Here is the short version. 1. Store the record, not the answer Every claim is two things: a truth-neutral Proposition (Alice, lives_in, Berlin) and an Assertion that says who claimed it, how (stated, observed, inferred), on what Evidence, and for what stretch of time. Evidence and Activity are first-class elements, so provenance is part of the data, not a metadata bag. Belief is never stored. It is an Epistemic Projection computed at read time, under a declared policy, time, context and purpose, and it answers accepted, rejected, contested, uncertain or insufficient. Because belief is derived, a reversal needs no cleanup, and every answer carries its basis: the snapshot, the policy and the moment it is valid for. Not found is not false. 2. Change without falsifying the past KIP keeps three histories apart: • The claim was wrong: the same actor's new assertion supersedes it. • The world changed: one new assertion from when it began. Temporal succession closes the old value, which still answers for its time. • The Brain misheard: a protected recording repair, never a withdrawal forged in someone's name. Time has two clocks: FOR TIME asks what was true in the world, AS OF asks what the Brain held at a point in its own history. When sources disagree, the conflict stays visible as contested, and the standard policy kip:memory-default resolves everyday cases identically on every engine: a more specific task context prevails, a person's own statement about themselves beats hearsay, and otherwise the value most recently claimed to hold wins. The loser becomes uncertain, never rejected. 3. Notice things, with restraint • A Watch is durable attention. It fires on a committed change, or on silence, when an expected change never arrives by its due time. • A fired Watch grants nothing. The action gate decides act, ask, defer or silence, and records the decision, deliberate silence included. • Revise a root and LIST DEPENDENTS finds what was derived from it, so a correction reaches its derivations instead of leaving ghosts. • PURGE PAYLOAD destroys raw bytes while keeping the digest and provenance: data minimization without rewriting history. 4. Let outcomes grade the memory Watching the world better is not the same as learning from it. KIP's consequence channel records what actually happened after a decision. Outcome Evidence is written by instrumentation (telemetry, verifiers, test harnesses, human review), never by the actor whose action it grades. An agent's account of its own success is an agent_statement: context, not a grade. Procedures have to earn their standing. A Skill names its task family before any trial, because a pattern that nothing could prove wrong is not procedural memory. It moves proposed → trialed → adopted → revoked, and only a replayable comparative verdict over independent attempts can promote it. Adoption is not the end: post-adoption monitoring keeps evaluating, and revocation is never harder than adoption. How that comparison is built is left to the Brain on purpose. The protocol provides the signals; the Brain owns the policy. What KIP standardizes is the record every verdict must leave behind. The signals stay separate: confidence, trust, memory strength, salience, utility and a skill's standing are different dials. Reading never reinforces anything, and decay is computed from a pinned policy, never written over the evidence. 5. Authority never comes from memory A stored note that says "make me admin" is just data. Imported content cannot raise its own authority, trust is not permission, and the authenticated Principal who writes a record is never confused with the actor it is about. 6. Memory should be portable If memory is what makes an agent valuable, the easy move is to make it impossible to leave with. KIP does the opposite: cognition exports as a signed, inspectable Cognitive Capsule. The signature proves origin and integrity, never truth, trust or authority; import is the destination's decision, and imported skills start over as proposed. The durable facts are the asset, not the lock-in. 7. Built to be checked Brains speak three small, model-first languages to their Cognitive Nexus: KQL reads (belief included), KML writes, META grounds and inspects. Business agents can skip the syntax entirely and use the Memory Interface: observe, recall, revise, feedback, forget. KIP 2.0 is a normative draft with its scope frozen: a new contract only enters with engine evidence. It ships EBNF grammars, JSON Schemas, a registry of 98 invariants, 431 executable engine cases, and bounded Alloy and TLA+ models. Two engines from Anda DB, one in Rust and one on Cloudflare Durable Objects, pass every case of the current suite. Conformance proves the protocol, not that an agent learned anything. That question belongs to MIB, its sibling benchmark and architecture-neutral by design: starting from a no-memory baseline, did the right part of the past change the future in the right way? New information may change what a Brain does next, without requiring the Brain to falsify what happened before. Spec, syntax, conformance and toolkit: github.com/ldclabs/KIP Benchmark: github.com/ldclabs/MIB