Agents that remember. Pipelines that can prove it.

Kumiho is the memory-and-provenance substrate for the agentic era — one versioned graph that gives AI agents auditable memory and gives creative teams version control for everything they make.

Start Free Watch a Demo

AI Cognitive Memory

A cognitive memory core, not an embedding cache.

Every memory is a typed, versioned node in a provenance graph. Recall survives context resets and model swaps — and every answer can show where it came from.

Belief revision, proven: AGM-compliant revise/contract/expand — 49/49 compliance scenarios pass against the shipped engine.
Typed provenance edges: DERIVED_FROM, DEPENDS_ON, REFERENCED — every memory knows its sources and its dependents.
Dream State: Async consolidation reviews, links, and prunes memory in the background, with safety guards.
Any agent, one MCP server: engage before responding, reflect after — plugs into Claude Code, Codex, or your own stack.

from kumiho import memory_store, memory_recall

# Store a decision with provenance
memory_store(
  summary="Approved hero layout",
  edges=[{"type": "DERIVED_FROM",
          "target": "kref://project/brief/rev/2"}]
)

# Recall by meaning — with sources attached
results = memory_recall("hero layout decision")
print(results[0].summary, results[0].edges)

Explore AI Cognitive Memory

Revka

Memory-native AI agent runtime.

An open-source (Apache-2.0) Rust runtime where every plan, tool call, and trust score lands in the same graph — with a Merkle-chained audit trail and declarative YAML workflows.

Drives the CLIs you already use

Claude Code
Codex
Antigravity
Cursor
OpenCode

Explore Revka GitHub

The foundation: version control for creative work

Agent memory is only as trustworthy as the substrate underneath. Kumiho’s is a production-grade provenance graph that started life tracking creative pipelines — and still does.

creatorsEvery prompt, model, and seed — kept: Generate in ComfyUI, iterate freely, and find “the one” weeks later by meaning, not filenames. Files stay on your machine or NAS.Creator workflows
studiosLineage and impact analysis across the pipeline: Connect plates → comps → finals with typed edges. Immutable publishes, audit trails, and blast-radius answers before a change ships.Studio pipelines
developersA graph API you don’t have to build: Revisions, edges, search, and events via Python, C++, and Dart SDKs — the same primitives that power agent memory.Developer APIs

Four primitives, every use case

The same building blocks underneath agent memory and studio pipelines. Nothing scale-dependent, nothing bolted on.

Immutable revisions: Every change becomes a revision you can tag, audit, and reference. History is never rewritten.
Typed edges: DERIVED_FROM, DEPENDS_ON, and friends connect revisions into a lineage you can traverse.
Bring-your-own storage: Artifacts are tracked by reference. Files stay on disk, NAS, or object storage — no uploads.
Events for automation: Every write emits an event. Drive render farms, publishing, and agent workflows without polling.

Meet it where you work: Kumiho Browser · ComfyUI · n8n · Revka · MCP tools · SDKs

See it run

A local-first walkthrough of the revision ledger, lineage view, and memory graph — the product surfaces are the pitch.

Watch the Product Demo Read the Docs

Ready to see Kumiho on your pipeline?

Tell us what you are building — agent memory, studio pipelines, creator workflows, or developer tooling — and we will help you map Kumiho into it.

Start Free Talk to Us