🧬 KIP (Knowledge Interaction Protocol)

English | 中文

The open experience learning protocol for AI agents —
turning interaction into memory, knowledge, skill, and better action.

KIP 2.0 is in draft — start here → Belief becomes an explicit projection over competing assertions instead of a stored field, evidence and provenance become first-class elements, and a proactivity layer lets a Brain act on a state differential rather than a timer: Watch, action_gate, LIST DEPENDENTS, PURGE PAYLOAD. v1.0-RC11 below remains the shipping contract; two independent engines — Rust and Cloudflare Durable Objects — already track the draft.

Why KIP?

An intelligence that cannot accumulate experience cannot truly learn.

Today's AI can reason brilliantly inside one session and still repeat the same mistake tomorrow. Larger context windows postpone forgetting; vector retrieval can return related text. Neither, by itself, explains what the agent was trying to achieve, which action changed the situation, where reality contradicted expectation, or how the past should change the next action.

KIP is an open protocol for the whole learning loop:

Experience → Memory → Knowledge → Skill → Action
     ▲                                      │
     └────────────── new feedback ──────────┘

• Experience preserves a subject's goal-directed state–decision–action–feedback trajectory.
• Memory lets past state participate in future computation.
• Knowledge compresses stable regularities from evidence and experience.
• Skill compiles experience into an action-selecting policy.
• Action applies that policy to the world and creates new experience.

The protocol connects two complementary kinds of machine intelligence:

• the LLM — a powerful but stateless probabilistic reasoning and policy engine;
• the Cognitive Nexus — a persistent, precise, auditable symbolic substrate for memory and learning.

The model interprets and acts; the graph preserves and reorganizes what matters; KIP is the language through which the past changes the future. It is not a database driver. It provides cognitive primitives for remembering, replaying, associating, reinforcing, correcting, consolidating, compiling skills, and forgetting.

What this gives you

• 🧭 Trajectory memory, not just transcripts — goals, actions, observations, outcomes, and prediction errors remain queryable as a coherent Experience.
• 🧠 Memory that survives the session — Events, Experiences, facts, preferences, insights, skills, and commitments live in a graph the agent can revisit.
• 🛠️ Learning without retraining — repeated successes and failures can update Knowledge and Skill in seconds, without a weight update.
• 🎯 Action-aware recall — an Action Briefing can return applicable Skills, analogous Experiences, constraints, risks, and commitments before the agent acts.
• 🔍 Auditable cognition — assertions carry provenance, author, confidence, and temporal state; derived Knowledge and Skills point back to their evidence.
• 🤖 A self that persists — $self can retain identity, values, lessons, commitments, and a history of changed behavior.
• 📦 Portable learning — idempotent Knowledge Capsules can back up, migrate, and exchange a self-describing memory graph.

KIP in 60 Seconds

The Cognitive Nexus is a graph. Concept Nodes are the things worth remembering; Proposition Links are typed assertions connecting them. Assertions can themselves be subjects or objects, so KIP can represent provenance, attribution, disagreement, and evolving belief.

The LLM operates the graph through three compact instruction sets:

Remember a fact with provenance:

UPSERT {
  CONCEPT ?dark_mode {
    {type: "Preference", name: "Dark Mode"}
    SET ATTRIBUTES { description: "Prefers dark UI themes in all apps" }
  }
  CONCEPT ?alice {
    {type: "Person", name: "Alice"}
    SET PROPOSITIONS { ("prefers", ?dark_mode) }
  }
}
WITH METADATA {
  source: "conversation:2026-06-11",
  author: "$self",
  confidence: 0.95,
  memory_strength: 0.80
}

Recall the strongest current assertions:

FIND(?pref.name, ?link.metadata.confidence, ?link.metadata.memory_strength)
WHERE {
  ?alice {type: "Person", name: "Alice"}
  ?link (?alice, "prefers", ?pref)
  FILTER(IS_NULL(?link.metadata.superseded) || ?link.metadata.superseded == false)
}
ORDER BY ?link.metadata.memory_strength DESC, ?link.metadata.confidence DESC
LIMIT 10

Associate without knowing the schema in advance:

FIND(?predicate, ?neighbor)
WHERE {
  ?link ({type: "Person", name: "Alice"}, ?predicate, ?neighbor)
}
LIMIT 50

When an agent wakes inside a graph it has never seen, DESCRIBE PRIMER tells it who it is, which domains exist, and which types and predicates it can use. The graph describes itself.

The Experience Learning Profile

KIP Core remains a general graph protocol. Experience learning is an additive, self-described Cognitive Memory Profile built with ordinary KIP capsules. No KQL, KML, or META syntax changes are required.

The model

An Event and an Experience may refer to the same real-world interval, but they are not interchangeable. Event is observer-oriented; Experience is subject-oriented. A conversation with no meaningful goal/action/feedback dynamics should remain an Event. A deployment attempt with hypotheses, tool actions, failures, revised state, and a terminal outcome should be an Experience.

Memory, Knowledge, and Action name functional roles in the learning loop, not mandatory universal Concept Types. Domain capsules define concrete semantic types; the Experience Learning Profile adds three Concept Types (Experience, ExperienceStep, Skill) and four Proposition Types (has_step, caused_by, derived_insight, compiled_to).

graph LR
    X["Experience"] -->|"has_step"| S1["ExperienceStep 0<br/>observation"]
    X -->|"has_step"| S2["ExperienceStep 1<br/>action"]
    X -->|"has_step"| S3["ExperienceStep 2<br/>feedback"]
    S3 -->|"caused_by"| S2
    X -->|"consolidated_to"| K["Knowledge"]
    X -->|"derived_insight"| I["Insight"]
    X -->|"compiled_to"| P["Skill"]
    P -->|"conditions"| A["Future action"]
    A -->|"creates"| NX["New Experience"]

ExperienceStep.index establishes temporal order. caused_by is optional and explicit: earlier does not mean causal.

Expectation is a learning signal

Experience becomes especially valuable where the world violates the subject's model:

expected observation → actual observation → prediction error → policy update

ExperienceStep therefore supports expected_observation, actual_observation, and prediction_error. The parent Experience may carry an aggregate surprise_score, which contributes to salience and consolidation priority.

Confidence is not memory strength

The profile separates three independent signals:

A true but rarely useful fact may retain high confidence while its memory strength falls. A vivid new Experience may have high salience and memory strength while its causal interpretation remains low-confidence. Maintenance must never use confidence as a proxy for retrieval frequency or forgetting.

From Experience to Skill

Procedural consolidation compares trajectories instead of summarizing a single transcript: