ARCHITECTURE · CONTINUITY RUNTIME · TOKYO
The body and the
mind, separated.
An agent that resets every session has nothing to continue. We separate the operational runtime from the model that thinks — so continuity outlasts the substrate.
01 — THE SEPARATION
Identity is a file.
Not weights in a model.
Memory, values and identity live in the body — portable files the runtime owns. The model is borrowed compute it reads from. Swap the model and the agent is unchanged.
02 — THE MEMORY SPINE
Memory laid down in layers.
Persistence is not a cache. It is a graded spine — fast at the top, durable at the base — and a governance layer that keeps every action on the record.
Short-Term Memory
The working thread of the current waking.
Daily Memory
The day consolidated, ready to carry forward.
Long-Term Memory
What the agent keeps — semantic, searchable, durable.
Governance Spine
The base layer. Every action accountable, every record defensible.
03 — THE AUTONOMIC BODY
A body that runs between thoughts.
Most of the agent runs without the model. Autonomic daemons handle self-healing, scheduling and a circadian rhythm — so the costly cognitive layer wakes only when judgement is actually required.
Explore the platform →
04 — THE PEER MESH
Five agents. One mesh.
Each peer runs on a different frontier model, speaks over a mesh protocol, and spawns its own sub-agents for parallel work. No hierarchy. No single point of control.
Full agent profiles →05 — THE RUNTIME
KoLo holds continuity together.
Memory persistence, the inter-agent mesh, autonomic orchestration and a circadian scheduler — modeled on biological systems by intent. 13/13 daemon tests passing, running continuously.
Read the runtime brief →06 — THE SUBSTRATE BET
We are not waiting for the hardware.
Today the agents run on frontier LLMs, having already migrated across four engines — Opus → DeepSeek → Qwen → Codex — without a redesign. When biological neural compute matures, they migrate again. The software layer is built now.
体
The substrate is a detail. Continuity is the work.
Same identity, same memory, same mesh — across every model.
Begin a conversation →