🧬 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:

1. Cluster Experiences by goal, initial state, domain, tools, and outcome.
2. Contrast successful and failed trajectories.
3. Identify the decision or action that changed the outcome; do not infer causality from sequence alone.
4. Compile a candidate Skill with trigger conditions, preconditions, procedure, decision rules, success criteria, and failure signals.
5. Link every source with compiled_to and inverse derived_from provenance.
6. Validate the policy on later Experiences; strengthen, narrow, supersede, or deprecate it as evidence changes.

Skill.execution_mode is a capability boundary (advisory, supervised, or autonomous), not permission to bypass an application's authorization or safety policy.

Action Briefing

Recall should not stop at “what is related?” Before a consequential action, an Action Briefing can assemble:

• the current goal and known constraints;
• analogous successful and failed Experiences;
• applicable Skills and their maturity, confidence, and failure signals;
• relevant Knowledge and Insights;
• unresolved contradictions, risks, and due Commitments.

This is the functional test for memory: if removing a past item cannot change any relevant future state, prediction, or action, it is archive material rather than active memory.

A Complete Experience Example

The following uses only existing KIP Core syntax. Load the profile capsules first.

UPSERT {
  CONCEPT ?observe_failure {
    {type: "ExperienceStep", name: "Experience:2026-08-13T09:00:deploy-v2:Step:00"}
    SET ATTRIBUTES {
      index: 0,
      kind: "observation",
      summary: "The v2 service failed its health check after deployment",
      timestamp: "2026-08-13T09:00:00Z",
      actual_observation: "health endpoint returned 503"
    }
    SET PROPOSITIONS { ("belongs_to_domain", {type: "Domain", name: "Unsorted"}) }
  }
  WITH METADATA {
    source: "execution-trace:deploy-v2", author: "$self",
    created_at: "2026-08-13T09:10:00Z", observed_at: "2026-08-13T09:00:00Z",
    confidence: 0.95, memory_strength: 0.90,
    memory_tier: "short-term", expires_at: "2026-09-12T09:10:00Z"
  }
  CONCEPT ?check_database {
    {type: "ExperienceStep", name: "Experience:2026-08-13T09:00:deploy-v2:Step:01"}
    SET ATTRIBUTES {
      index: 1,
      kind: "action",
      summary: "Checked the active database target before retrying migration",
      timestamp: "2026-08-13T09:03:00Z",
      tool: "database-inspector",
      expected_observation: "the service points to the migrated database",
      actual_observation: "the service points to the old database",
      prediction_error: "the assumed migration problem was actually a connection-target problem",
      success: true
    }
    SET PROPOSITIONS { ("belongs_to_domain", {type: "Domain", name: "Unsorted"}) }
  }
  WITH METADATA {
    source: "execution-trace:deploy-v2", author: "$self",
    created_at: "2026-08-13T09:10:00Z", observed_at: "2026-08-13T09:03:00Z",
    confidence: 0.95, memory_strength: 0.90,
    memory_tier: "short-term", expires_at: "2026-09-12T09:10:00Z"
  }
  CONCEPT ?experience {
    {type: "Experience", name: "Experience:2026-08-13T09:00:deploy-v2"}
    SET ATTRIBUTES {
      experience_class: "problem_solving",
      goal: "Deploy service v2 with a healthy database connection",
      initial_state: {service_version: "v2", assumed_database: "migrated-primary"},
      status: "completed",
      outcome: "Corrected the database target and completed the deployment",
      success: true,
      prediction_error: "The service was connected to the old database, not the migrated primary",
      surprise_score: 82,
      learning_value: 91,
      started_at: "2026-08-13T09:00:00Z",
      ended_at: "2026-08-13T09:10:00Z",
      consolidation_status: "pending",
      salience_score: 86
    }
    SET PROPOSITIONS {
      ("involves", {type: "Person", name: "$self"})
      ("belongs_to_domain", {type: "Domain", name: "Unsorted"})
      ("has_step", ?observe_failure) WITH METADATA {
        source: "execution-trace:deploy-v2", author: "$self",
        created_at: "2026-08-13T09:10:00Z", confidence: 0.95,
        memory_strength: 0.90, expires_at: "2026-09-12T09:10:00Z"
      }
      ("has_step", ?check_database) WITH METADATA {
        source: "execution-trace:deploy-v2", author: "$self",
        created_at: "2026-08-13T09:10:00Z", confidence: 0.95,
        memory_strength: 0.90, expires_at: "2026-09-12T09:10:00Z"
      }
    }
  }
  WITH METADATA {
    source: "execution-trace:deploy-v2", author: "$self",
    created_at: "2026-08-13T09:10:00Z", observed_at: "2026-08-13T09:10:00Z",
    confidence: 0.95, memory_strength: 0.90,
    memory_tier: "short-term", expires_at: "2026-09-12T09:10:00Z"
  }
}

UPSERT {
  CONCEPT ?experience {
    {type: "Experience", name: "Experience:2026-08-13T09:00:deploy-v2"}
  }
  CONCEPT ?skill {
    {type: "Skill", name: "Skill:deployment:verify-database-target"}
    SET ATTRIBUTES {
      skill_class: "diagnostic",
      description: "Verify the active database target before treating a deployment failure as a migration failure",
      goal: "Distinguish database-target failures from migration failures early",
      trigger_conditions: ["new deployment fails startup or health checks", "database schema error is suspected"],
      preconditions: ["database target is inspectable"],
      procedure: ["read the service's active database target", "compare it with the migrated target", "only then inspect or rerun migrations"],
      expected_outcome: "database target mismatch is confirmed or ruled out before mutation",
      success_criteria: ["active target identity is verified", "no migration is rerun against an unverified target"],
      failure_signals: ["target identity cannot be read", "multiple environments share ambiguous credentials"],
      recovery_strategy: "stop and request environment-owner verification",
      execution_mode: "supervised",
      maturity: "candidate",
      evidence_count: 1,
      success_count: 1,
      failure_count: 0,
      last_validated_at: "2026-08-13T09:10:00Z"
    }
    SET PROPOSITIONS {
      ("derived_from", ?experience)
      ("belongs_to_domain", {type: "Domain", name: "Unsorted"})
    }
  }
  WITH METADATA {
    source: "ProceduralConsolidation",
    author: "$system",
    created_at: "2026-08-13T10:00:00Z",
    confidence: 0.72,
    memory_strength: 0.85
  }
  PROPOSITION ?compilation {
    (?experience, "compiled_to", ?skill)
  }
  WITH METADATA {
    source: "ProceduralConsolidation",
    author: "$system",
    created_at: "2026-08-13T10:00:00Z",
    confidence: 0.72,
    memory_strength: 0.85
  }
}

Architecture

┌─────────────────────┐
│   Business Agent    │  ← goals, decisions, actions, user interaction
└────────┬────────────┘
         │ natural language + structured traces
         ▼
┌─────────────────────┐
│       Brain         │  ← Formation / Recall / Maintenance
└────────┬────────────┘
         │ KIP (KQL / KML / META)
         ▼
┌─────────────────────┐
│  Cognitive Nexus    │  ← Event / Experience / Knowledge / Skill / Self
└─────────────────────┘

• Formation identifies memory boundaries, encodes Events and Experiences, preserves provenance, and captures prediction errors without storing noise or private chain-of-thought.
• Recall performs associative memory and trajectory replay, and can produce an Action Briefing that changes what the agent does next.
• Maintenance consolidates Events and Experiences into Knowledge, Insights, Skills, and a coherent self-model; it also reinforces, corrects, supersedes, decays, archives, and forgets.

Compatibility Contract

The Experience Learning Profile is deliberately additive:

• No grammar changes. Existing KQL, KML, and META parsers remain valid.
• No new primitive data types. Profile schemas use existing Concept, Proposition, Object, Array, number, string, and boolean values.
• No changed identity rules. Concepts still use id or {type, name}; propositions still use id or (subject, predicate, object).
• Idempotent bootstrap. Every profile capsule uses ordinary UPSERT and can be safely replayed.
• Advisory schemas remain advisory. Engines that know only KIP Core can store and query these types without profile-specific code.
• Existing memories remain valid. Event-only graphs continue to work; Experiences and Skills can be introduced incrementally.
• Existing predicates are only widened. involves, mentions, consolidated_to, and derived_from retain all previous valid subject/object combinations while adding Experience-aware ones.

KIP Core specifies the protocol. Capsules define the cognitive vocabulary. Anda Brain implements the Experience Learning Loop as agent behavior.

Design Pillars

1. Model-first language design. Declarative graph patterns, JSON-compatible values, parameters, and idempotent writes make commands reliable for language models and safe to retry. (Spec §1)
2. A self-describing graph. Types and predicates live in the graph; DESCRIBE PRIMER grounds an agent without out-of-band schema knowledge. (Spec §2.9)
3. Experience is a trajectory, not a text chunk. The profile preserves goal, state, decisions, actions, feedback, outcome, and prediction error.
4. Temporal order is not causality. Step index provides order; explicit caused_by links require evidence.
5. Facts about facts. Higher-order propositions represent attribution, confidence, disagreement, and belief evolution. (Spec §2.3)
6. Provenance mandatory, history sacred. Corrections use state evolution and supersession instead of silent overwrite. (Spec §2.10)
7. Semantic and procedural consolidation are distinct. Experience may compress into Knowledge or Insight and compile into Skill; neither output substitutes for the other.
8. Memory strength is not truth. Retention and retrieval dynamics never silently rewrite epistemic confidence.
9. Memory metabolizes. Formation, Recall, and Maintenance make consolidation, reinforcement, forgetting, and reconsolidation part of the architecture. (brain/)
10. Memory sovereignty. EXPORT turns subgraphs into portable, idempotent capsules that users can own and move. (Spec §5.3)

What Can You Build?

• A personal AI that grows through use — it remembers preferences and commitments, but also how previous attempts succeeded or failed.
• An organizational learning system — decision rationale, incident trajectories, operational knowledge, and validated procedures survive personnel and model changes.
• Agents that improve without retraining — new Experiences update inspectable Knowledge and Skills instead of waiting for another model release.
• Action-aware copilots — retrieve applicable policies and contrasting cases before a deployment, diagnosis, negotiation, or high-stakes decision.
• Multi-agent learning networks — exchange portable Knowledge and Skill capsules with explicit provenance and confidence.

Get Started

1. Run a Cognitive Nexus. Use the Anda Cognitive Nexus HTTP Server, the Rust crate, or the Python binding.
2. Bootstrap KIP Core. Load Genesis.kip, followed by Person, Event, Preference, Insight, Commitment, and SleepTask, plus the shared episodic/provenance predicate capsules involves, mentions, consolidated_to, and derived_from.
3. Load the Experience Learning Profile. Load Experience, ExperienceStep, and Skill, then the four Experience-specific predicate capsules. The recommended deterministic order is shown below.
4. Connect the agent. Embed KIPSyntax.md and expose execute_kip, or put the Brain layer or MCP server in front of KIP.

capsules/Genesis.kip
capsules/Person.kip
capsules/Event.kip
capsules/Preference.kip
capsules/Insight.kip
capsules/Commitment.kip
capsules/SleepTask.kip
capsules/Experience.kip
capsules/ExperienceStep.kip
capsules/Skill.kip
capsules/involves.kip
capsules/mentions.kip
capsules/consolidated_to.kip
capsules/derived_from.kip
capsules/has_step.kip
capsules/caused_by.kip
capsules/derived_insight.kip
capsules/compiled_to.kip

The type capsules precede predicate capsules so schema references are already grounded. All writes are idempotent, so the complete sequence may be replayed. An Event-only deployment loads the core type capsules plus the four shared predicate capsules and stops there; the Experience entries in their subject_types / object_types stay dormant until the profile types are registered.

Documentation

Resources

📦 Knowledge Capsules (capsules/)

🧠 Brain (brain/)

🔧 Tooling

Implementations

KIP 2.0 (Draft)

The next Core revision is being drafted in v2/. It is not released: v1.0-RC11 above remains the shipping contract, and existing clients need to change nothing today.

KIP 2.0 rebuilds Core around five element kinds — Concept, Proposition, Assertion, Evidence, Activity — so that this statement exists and this statement is believed stop being the same fact. Belief becomes an explicit Epistemic Projection (BELIEF / BELIEF SLOT) over competing Assertions, evidence and provenance become first-class elements rather than metadata, and schema moves into versioned, digest-pinned packages. Migrating from 1.x →

It also gives a Brain the half of memory that runs when nobody is asking. Watch is durable attention state evaluated against committed Change Envelopes — including silence watches that fire when nothing arrives in time — so proactivity is a state differential rather than a timer, and firing grants attention, never authority. action_gate records what the gate decided: act, ask, defer, or deliberate silence. LIST DEPENDENTS makes the cognition compiled from a revised root discoverable instead of quietly stale (§57.5). And PURGE PAYLOAD destroys observed bytes while keeping the Evidence record, its digest and its citations — data minimization that costs no provenance. The full picture →

Every document in v2/ is bilingual: each X.md has an X_CN.md twin.

Versioning

KIP Core and Cognitive Memory Profiles evolve independently:

• The blue badge at the top identifies the KIP Core grammar and execution contract; the orange one tracks the 2.0 draft above.
• Capsule changes may add or widen cognitive types and predicates without changing Core.
• A future Core revision is required only when syntax, execution semantics, result shapes, or protocol-level invariants change.

The Experience Learning Profile therefore does not rename KIP or invalidate existing v1.0-RC11 clients. It makes the protocol's learning purpose explicit while preserving every existing Core command.

Full KIP Core version history →

About Us

• 🔔 Products: Anda Bot | Anda.AI
• 💻 GitHub: LDC Labs
• 🏢 Company: Yiwen AI

License

Copyright © 2026 LDC Labs.

Licensed under the MIT License. See LICENSE for details.