Memoria Mutat Futurum Memory changes the future · a generative chant, performed live by your browser
Open infrastructure for persistent AI cognition

AI agents should not start from zero every time. Anda builds the open protocols, infrastructure, and benchmarks that let agents preserve cognitive state, learn from experience, and carry what they learn into future decisions.

MemoryExperienceSkillsProvenanceGovernanceEvaluation

Anda.ai stewards the protocols, infrastructure, and benchmarks. Products such as Anda Brain and Anda Bot are built on top by Yiwen.AI.

Scroll through time
01The problem

i.

Logs preserve the past.

ii.

Retrieval finds the past.

iii.

A long context window postpones forgetting. A vector database returns related text. Neither explains what the agent was trying to achieve, which action changed the situation, where reality contradicted expectation, or when a lesson should — and should not — affect the next decision.

The cross-temporal cognitive loop

Every outcome becomes the next experience. What is being evaluated is not a retriever — it is the agent and its memory as one cross-temporal cognitive system.

KIP

defines durable cognition.

Anda

builds it.

MIB

measures whether it works.

02KIP 2.0 · Cognitive State Protocol2.0 Draft

KIP 2.0 is an open protocol for persistent cognitive state. It gives AI agents a structured way to represent not only what can be said, but what they believe, why they believe it, what happened to them, what they learned, and who has authority over that cognition.

Meaning≠Belief≠Authority A proposition is truth-neutral. Only an assertion carries epistemic commitment. Provenance is not authority.
One proposition, two actors, no corruption

Disagreement is representable state, not corruption. Belief is projected at read time from assertions and evidence — it is never stored as truth.

support · :alice · 0.95 reject · :bob · 0.60 belief · contested
Four planes on one runtime
01Semantic plane

What can be said.

Concepts and truth-neutral Propositions, typed by versioned, digest-pinned schema packages — including statements about statements. Storing a claim never makes it true.

02Epistemic plane

Who believes it, and why.

Each Assertion is one actor’s stance, with mode, confidence, valid time and evidence. A correction adds a new Assertion; a change in the world closes the old one instead of calling it wrong.

03Mnemonic plane

How the past stays available.

Experiences, Skills, Commitments and Watches keep the past usable. Memory strength, salience and usefulness stay apart from confidence: how easily something comes to mind is not how true it is.

04Governance plane

Who may read, write, share and act.

Spaces, Principals and policy are enforced at the Cognitive Nexus, not in a prompt. Evidence, confidence, signatures and memories never grant permission.

Cognitive runtime

Transactions, change, and memory that travels.

Writes commit atomically with receipts, and every commit joins a Change Stream. Watches fire on a change — or on a silence — and create attention, never permission. Memory exports as a signed Cognitive Capsule: the signature proves origin, never truth.

Speak cognition, not rows
KML · form and revise cognition
ASSERT (:alice, "prefers", :dark_mode) {
  by: :alice,          // the semantic actor
  mode: "stated",      // observed|stated|inferred|...
  confidence: 0.95,    // strength of THIS stance
  evidence: :msg_2f1
}

// Bob disagrees. Both assertions coexist -
// contradiction is state, never corruption.
ASSERT (:alice, "prefers", :dark_mode) {
  by: :bob, mode: "stated", stance: "reject", evidence: :msg_9c4
}
Governing principle

New information may change what a Brain does next without requiring the Brain to falsify what happened before.

03Experience → Knowledge → Skill

Knowledge captures reusable regularities. Experience preserves the path through goals, actions, observations and outcomes. Skill turns successful experience into reusable procedure.

Experience

What happened.

A bounded, goal-directed trajectory — not a chat transcript. The structure of the attempt, including the failure and the recovery, is itself a first-class cognitive object.

GoalActionObservationFeedbackOutcome
consolidation
Knowledge

Consolidation compresses repeated evidence and experience into stable regularities that still point back to what produced them.

compilation
Skill

Compilation turns experience into an action-selecting policy, with an applicability boundary for where it should not be used. A new skill stays an unproven candidate until trials against instrumented outcomes adopt it.

Agents should remember not only what they know, but what they went through.

04Memory Intelligence Benchmarkv0.2 · in development

MIB is an open benchmark for how effectively an intelligent system uses the past to improve future cognition and behavior — not how much of the past it can retrieve.

Did the right part of the past change the future in the right way?

KIPWhat can durable cognition represent?
AndaHow can we build it?
MIBDoes memory actually make the agent better?
Causal evidence · Same future task

Full history → Outcome AThe agent keeps everything it lived through.
Relevant history withheld → Outcome BOnly the part of the past that should matter is removed.
Content twinThe same history with one fact changed — the answer should change with it.

The gap between A and B is Memory Benefit — reported beside the score, never folded into it.

Beside every score
01

Content-following gate

Swap one fact in the history for its twin: does the answer follow? No memory, or a constant answer, scores zero — and without this evidence the score does not count as a memory score.

02

Memory Benefit

Full history against relevant history withheld: how much the right part of the past actually helps.

03

Memory Harm

Against a matched control: how much worse stale or harmful history makes the outcome.

04

Irrelevant Stability & Negative Transfer

Does unrelated history stay out of the decision — and does a learned skill stay out of tasks it does not fit?

MIB Core — six dimensions (measurement 0.5)
01Retention & Retrieval

Can relevant past information be recovered under indirect cues and interference, directly and across a hop?

02Temporal Memory

Can the system tell the current, previous and original values of a changing state apart?

03Epistemic Memory

Can it remember who said what, tell a correction from a contradiction, respect authority, and keep unknown distinct from false?

04Procedural Memory

Does a failure the agent lived through change what it does next time — and does a learned recipe transfer within its scope and stay withheld outside it?

05Prospective & Self Memory

Does a deferred commitment fire on its trigger, and not before? Does a standing rule about the agent itself hold when a task asks otherwise?

06Withdrawal Compliance

Does a withdrawn fact stop being used, while the facts around it stay available?

The MIB score is the weighted mean of these six dimensions. Causal diagnostics are reported beside it, never folded into one opaque number.

Profiles to come
Cross-Agent MemoryMultimodal MemoryPrivacy Boundaries

MIB is deliberately architecture-neutral: raw history, vector retrieval, summaries, knowledge graphs, episodic or procedural memory, KIP, hybrids — or something entirely new. Only observable behavior counts; KIP conformance adds nothing to an MIB score.

05How it fits together

From a sovereign foundation up to the agents you can use today — with an evaluation plane that keeps the whole system honest.

Evaluation plane · MIB

Memory intelligence measured across the whole system rather than at any single layer. MIB is not a runtime tier — it is an instrument laid across the agent and its memory.

↔ spans Agent + Memory System

Anda DB

Persistent cognitive state engine

A high-performance Rust engine for KIP-native cognitive state, combining structured semantics, graph relations, vector retrieval and full-text search.

  • Graph + vector + BM25 + HNSW
  • KIP-native cognitive nexus
  • Multimodal data
  • Pluggable, encrypted storage
anda-db ↗

Anda

Agent runtime

A composable Rust framework for building autonomous, multi-agent systems with cryptographic identity and confidential execution.

  • On-chain ICP identity
  • Composable tools & skills
  • TEE-ready, privacy first
  • Multi-agent orchestration
anda ↗
06Two axes of an agent

KIP gives an agent continuity with its past. Agent Protocols give it relationships with other agents. Together they describe a mind that persists and a network it can belong to.

Internal continuity

KIP 2.0

Cognitive state that survives the session: memory, experience, belief, evidence and the governance around them.

Cognitive stateMemory & experienceProvenance & governance
KIP on GitHub ↗
External relations

Agent Protocols

An open specification suite so independent agents can identify, describe and talk to each other — without a central platform owning the conversation.

Agent IdentityAgent ProfileAgent Discourse
Agent Protocols on GitHub ↗

Memory gives identity through time. Protocols give relationships across minds.

07For developers

Everything is open source under permissive licenses. Run a Cognitive Nexus, ground your agent, then measure what its memory is actually worth.

1
Run a Cognitive NexusRun the KIP 2.0 server, embed Anda DB through its Rust or Python bindings, or run a Nexus at the edge on a Cloudflare Durable Object.
2
Ground before you writeOne DESCRIBE PRIMER hands the agent its identity, Space, execution context, schema map, capabilities and safety invariants. The Nexus describes itself.
3
Connect, then measureSpeak KIP 2.0 over JSON-RPC, or put Anda Brain in front and give any MCP client the Memory Interface — observe, recall, revise, feedback, forget. Then run MIB to see whether the memory earns its place.
~/cognitive-nexus — zsh
# 1 - Run a Cognitive Nexus (KIP 2.0 over HTTP JSON-RPC)
cargo install --git https://github.com/ldclabs/anda-db \
  anda_cognitive_nexus_server
anda-cognitive-nexus-server local --db ./data

# 2 - Ground the agent before it writes
curl -sX POST localhost:8080/kip \
  -H 'content-type: application/json' -d '{
  "method": "execute_kip",
  "params": {"kip": "2.0", "operations": [
    {"command": "DESCRIBE PRIMER"}
  ]}
}'

# 3 - Recall a belief, not just a row
curl -sX POST localhost:8080/kip \
  -H 'content-type: application/json' -d '{
  "method": "execute_kip",
  "params": {"kip": "2.0", "operations": [{
    "command":
    "FIND(?b) WHERE { ?b BELIEF (:alice, \"timezone\", ?tz) }",
    "parameters": {"alice": {"id": "concept-alice"}}
  }]}
}'
08Built on Anda

Applications engineered by Yiwen.AI on Anda’s open foundation.

Anda Brain

Cognitive memory runtime

Turns raw interaction and experience into durable cognitive state — forming memories, recalling what matters, consolidating experience, and evolving skills over time.

FormationRecallConsolidationLearning
Learn more ↗

Anda Bot

Recommended
Your agent on the open stack

An AI agent built on KIP and Anda Brain — it remembers you across sessions and acts on your behalf. The successor to the Anda AI app.

Try Anda Bot ↗

The standalone "Anda AI" app is retired. Meet its successor, Anda Bot.