KoLo Platform
The operating system for persistent AI agents.
KoLo is a continuity-first runtime for running, observing, governing, and moving AI agents across models, cloud, edge, HPC, and future biological-compute substrates.
Why agents need an OS
Agents need more than prompts. They need runtime infrastructure.
Most AI agents are still built as chat sessions, workflow chains, or model-specific services. KoLo provides the missing operating layer: memory, scheduling, permissions, telemetry, recovery, and substrate abstraction.
The model thinks. KoLo remembers, schedules, governs, and recovers.
OS primitives
Six primitives. One operating layer.
Memory filesystem
Persistent short-term, daily, and long-term memory. Markdown + JSON on disk. Diffable, versioned, recoverable.
Runtime scheduler
Background cycles, health checks, and maintenance windows. 144 autonomic cycles per agent per day at zero token cost.
Agent process model
Agents run as persistent supervised processes, not isolated prompt calls. State outlives any single inference request.
Mesh networking
Type-safe ordered A2A communication between agents. Peer-oriented routing with human-governed oversight.
Governance kernel
Permissions, audit logs, approval rules, and policy constraints. Replayable, inspectable, exportable.
Substrate drivers
LLMs today. Edge, HPC, neuromorphic, and biological-compute interfaces as they mature.
Deployment model
One request path. Substrate-portable.
Developer / Partner
↓
KoLo Gateway
↓
Agent Runtime
↓
Memory + Governance Layer
↓
Substrate Adapter
↓
LLM · Edge · HPC · Future BioComputeKoLo separates agent state from the inference substrate. This allows persistent agents to move across providers and deployment environments while preserving approved memory, runtime state, governance context, and audit history.
Proof & telemetry
Numbers, not narrative.
68:1
Autonomic-to-cognitive runtime ratio
144 / day
Daemon cycles per agent
139+
Memory versions preserved
4 / 4
Frontier-model transitions completed with continuity state preserved
13 / 13
Daemon module tests passing, continuous since March 2026
5-agent mesh
Active research topology across heterogeneous substrates
Telemetry samples are available to qualified partners under NDA. The April 17, 2026 security-stress event — five encrypted backup bundles generated by the daemon during degraded network and provider-instability conditions — is included as a reference case.
Technical specs
Boring, on purpose. Engineering, not narrative.
| Layer | Current implementation |
|---|---|
| Runtime core | KoLo Gateway · Node.js |
| Agent engine | Daemon V2 · Python |
| Memory format | Markdown + JSON · diffable, versioned, recoverable |
| Messaging | A2A mesh protocol · type-safe, ordered, transactional |
| Substrate adapter | Anthropic · DeepSeek · Qwen · OpenAI Codex (tested) |
| Continuity benchmark | 68:1 autonomic-to-cognitive · 139+ memory versions |
| Future work | Neural-substrate adapter spec drafted, awaiting hardware partner |
Build on KoLo.
The runtime is operational. The substrate is open.
We are looking for research, infrastructure, HPC, and biological-compute partners to test KoLo against real substrate constraints: latency, failure recovery, observability, memory portability, and governed autonomy.