We built provenance for studios. Agents turned out to need it more.
What slows creative teams down is not making new work — it is losing context: which version is correct, how it was made, and what depends on it. We built Kumiho to fix that with a graph-native system for revisions and relationships, while files stay on your own storage.
Then we watched AI agents hit the same wall, harder. They forget between sessions, cannot trace why a decision was made, and contradict themselves when memory is a pile of embeddings.
So Kumiho became the memory-and-provenance substrate for the agentic era: AI Memory with formal belief revision, and Revka, a memory-native agent runtime — both running on the versioned graph that studios already trusted with their history.
Who we are
We are a small team of creators and pipeline engineers. We spent years inside VFX, animation, and AI workflows where high-value work gets buried in folder chaos — and where "why did we do it this way?" has no answer.
Kumiho started as our attempt to make revision tracking and dependency intelligence available without the infrastructure burden. It grew into the substrate we now point at a harder problem: memory agents can prove.
Why teams trust us
- Published the formalism: Graph-Native Cognitive Memory for AI Agents (arXiv:2603.17244)
- Benchmarks run in public — the harness, results, and AGM compliance suite are committed to GitHub
- Built creative pipeline tooling at scale across VFX, animation, and AI workflows
- Shipped production software used by teams: plugins, SDKs, browser tooling, automation nodes
- Patent-pending architecture filed with the USPTO
The insight that connects both worlds
Not two products. One graph applied twice.
- the creative problemStudios lose history
- Thousands of outputs, no ground truth. Shipping with confidence takes immutable revisions, typed dependencies, and full provenance.
- the memory problemAgents lose context
- Decisions, preferences, and facts evolve. Without structure, memory is shallow recall that cannot explain itself — or notice it contradicts itself.
- the realizationSame primitives, both worlds
- Revisions preserve history. Typed edges preserve meaning. Formal belief revision keeps memory correct when the world changes.
Shared architecture primitives
| Primitive | AI memory | Creative assets |
|---|---|---|
| Immutable revisions | Belief history for agent decisions | Version history for renders and models |
| Typed dependencies | Evidence chains and supersession links | Impact analysis for upstream changes |
| BYO storage | Conversation artifacts stay local | Raw files stay on local/NAS/S3 |
| Hybrid retrieval | Recall by meaning, path, or exact text | Find by metadata, text, or relation |
Why now
Three shifts made this inevitable.
- Output explosion
- Agent interactions and creative generation both produce variants faster than anyone can manually track.
- Provenance is required
- AI teams and studios now need reproducibility, explainability, and audit trails as baseline requirements — not features.
- Automation needs structure
- Reliable triggers require explicit relationships. A graph with typed edges is what makes automation safe at scale.
When work becomes automated, your knowledge system has to become operational.
How we build
- Receipts over adjectives
- Every claim carries its number, citation, or link. If a proof artifact is not public yet, the claim waits.
- Your data, your storage
- Raw content stays in your infrastructure. Kumiho stores metadata, relationships, and pointers — and Community Edition never leaves your machine.
- Nothing silently changes
- Revisions are immutable. Updates happen through new revisions, never hidden overwrites.
- Relationships are explicit
- Vector similarity helps discovery, but typed graph edges encode why things are connected.
- Memory outlives the model
- Switch tools or LLM providers without losing accumulated context and provenance.
- Safety by default
- Consolidation runs behind guardrails: no silent mass deprecation, no unsafe mutation of published records.
What is real today
Shipping software, not slideware.
AI Memory
The cognitive memory core — MCP server, Claude Code plugin, and SDKs with belief revision and Dream State.
Revka
Memory-native AI agent runtime with a tamper-evident audit trail.
Kumiho Browser
Browse revisions, compare versions, and trace lineage from one interface.
ComfyUI plugin
Capture prompts, seeds, and generation metadata as structured revision history.
n8n nodes
Trigger downstream workflows when assets or memory records change.
SDKs
Python and more — build custom workflows and integrations on the graph primitives.
Give your work — and your agents — a memory.
Join the creators, studios, and AI builders who treat context and provenance as assets worth keeping.
Free tier available. No forced uploads. No lock-in.