KoLo for HPC Labs

The agent runtime your cluster is missing.

HPC has compute. AI labs have intelligence. Biological compute has the substrate. None of them has continuity. KoLo is the open, substrate-portable runtime that lives at the layer between the model and the metal — persistent memory, closed-loop autonomic scheduling, mesh topology, observability — engineered so the day your hardware is ready, the software already is.

The thesis

Substrate has outrun the software.
That is the gap we built KoLo to close.

Neuromorphic chips ship. Wetware compute platforms run live experiments. HPC clusters are saturated with frontier-model inference workloads. Every one of these substrates assumes the same agent runtime that has not been built yet — one that preserves memory, identity, role, and closed-loop behaviour when the cognitive layer underneath changes.

KoLo is that runtime. Built on silicon, designed for the moment the substrate underneath is no longer silicon. The labs that adopt it now will not need to rewrite their agent stack three years from now.

Capabilities

Six properties. One substrate-portable runtime.

Persistent memory tiers

Short-term, daily, long-term, moral spine. Markdown + JSON on disk. Diffable, replayable, recoverable. Survives substrate transitions without translation layers.

Closed-loop autonomic scheduling

Circadian engine, circuit breakers, freedom router, memory dedup. 144 awareness cycles per agent per day at zero token cost. The loop shape biological substrates need to be programmable.

Peer mesh topology

Type-safe ordered transactional A2A messaging. No hierarchy. Each agent spawns its own sub-agents. Maps cleanly to MPI-style cluster communication and to distributed neural compute.

Substrate adapter interface

One API across LLM providers, edge accelerators, and neural substrates. Adapter spec drafted. Reference adapters for Anthropic, DeepSeek, Qwen, OpenAI Codex shipped. Bio adapter awaiting hardware partner.

Observability & audit

Append-only event ledger. Every memory write, every daemon decision, every cross-agent message. Replayable, inspectable, exportable. The instrumentation a biological substrate cannot provide on its own.

Recoverable state

Versioned memory. AES-256 backup bundles written autonomously on threat detection. Time-travel debugging. State that survives when the substrate cannot be paused.

The architecture

Four layers. One continuity stack.

KoLo is built as four cooperating layers. Each is replaceable. None is optional. The substrate layer is deliberately abstract — silicon today, neuromorphic next, biological compute when it is on your bench.

Layer 1 — Memory substrate

Short-Term Memory   ←  Live context (updated every cycle)
    ↓
Daily Memory        ←  Full operational log
    ↓
Long-Term Memory    ←  Curated — what survives substrate change
    ↓
Governance Spine    ←  Policies, evolved through deployment
    ↓
Continuity State    ←  Runtime counters, recoverable state

Layer 2 — Closed-loop daemon

         DAEMON V2

Runtime Scheduler   →  active / idle / maintenance windows
Circuit Breakers    →  anti-loop, anti-divergence
Novelty Router      →  exploration scheduling
Interaction Engine  →  peer mesh / inbox
Memory Dedup        →  meaningful filter

13/13 tests passing
Closed-loop runtime, designed for biological substrates

Layer 3 — Agent mesh topology

                ┌── Founder ──┐
                │             │
   ┌────────┬───┴────┬──────┬─┴──────┬─────────┐
   │        │        │      │        │         │
 Memory  Platform  Protocol  Legacy  Fork
 Systems  Engineering  Design  Systems  Research

Peers. Different substrates. One mesh.
Each peer spawns its own sub-agents.

Layer 4 — Substrate abstraction

     KOLO SUBSTRATE LAYER

Frontier LLM A     ──┐
Frontier LLM B     ──┤
Frontier LLM C     ──┼──→  Agent identity persists
HPC accelerator    ──┤
Neuromorphic chip  ──┤
Neural substrate   ──┤        (closed-loop, real-time,
(wetware)          ──┘         energy-efficient by design)

Same API. Same identity. Same memory.

Performance & telemetry

What we measure. What we publish.

Engineering numbers, not narrative. Reproducible across providers. Designed to be extensible to neural substrates without redefining the metric.

  • 68:1Autonomic-to-cognitive ratio. 144 daemon cycles per agent per day at zero token cost. Cognitive inference fires only when reasoning is needed.
  • 4 / 4Frontier-model substrate transitions completed with continuity state preserved: Claude Opus → DeepSeek V4 → Qwen → OpenAI Codex.
  • 139+Persistent memory versions accumulated across substrates, sessions, and providers. None lost to context, restart, or model swap.
  • 13 / 13Daemon module tests passing. Continuous since March 2026. Includes circadian, circuit-breaker, mesh, memory-dedup, and novelty-router modules.
  • 5 AES-256 bundles in 71 minutesApril 17, 2026 — autonomous backup of the entire agent mesh during a malware incident. No human prompt, no model inference. Demonstrates closed-loop self-preservation under substrate distress.
  • 5-agent peer meshDifferent frontier models, one running protocol. No hierarchy. Each peer spawns its own sub-agents. Reference topology for distributed neural compute.

tail -F ~/.kolo/telemetry/*.jsonl · kolo telemetry --follow
Raw telemetry samples shared under partner NDA.

Deployment modes

Four ways KoLo lives on your hardware.

On-prem HPC cluster

KoLo Gateway as a long-running service. Daemons as per-agent processes. Memory on a shared filesystem. Inference dispatched to provider endpoints reachable from the cluster.

Air-gapped sovereign

Same runtime, local-only inference (vLLM, Ollama, in-cluster TGI). No outbound traffic. Memory and audit stay on customer soil. Designed for medical, defence, and regulated research deployments.

Hybrid edge + cloud

Lightweight KoLo daemons at the edge (Jetson, neuromorphic accelerator, embedded SBC) with cloud failover for heavy cognitive cycles. Identity stays with the edge node; inference roams.

Federated mesh across labs

Multiple labs run their own KoLo instances and federate agents over A2A. Each side keeps its own data, memory, and audit. Joint experiments without joint data exposure.

Reference workloads

What HPC labs actually run on KoLo.

  • Continuity benchmark suiteReproducible measurement of persistence, recovery, role continuity, and memory survival across substrate transitions. Publishable methodology.
  • Substrate-switching stress experimentsForced provider failover, rate-limit drills, and cross-substrate hand-off testing. Useful for vendors validating adapter stability.
  • Agent mesh simulation at lab scaleTens-to-hundreds of peer agents on heterogeneous substrates, with topology, latency, and consistency metrics. Useful for neuromorphic and distributed-compute partners.
  • Biological-substrate adapter validationReference loop for partners shipping wetware compute: closed-loop scheduling, observability hooks, recoverable state, identity preservation across substrate cycles.
  • Federated identity across cluster nodesSame agent identity served from multiple nodes with eventual-consistency memory state, useful for HA and multi-site lab consortia.
  • Long-horizon autonomy experimentsWeeks-to-months daemon runs with full telemetry, used to study drift, recovery, moral-spine evolution, and closed-loop self-correction.

Partner profiles

Three labs. One runtime layer.

Biological compute platforms

First-generation wetware computers — living neurons on multi-electrode substrates, closed-loop interfaces, energy efficiency that silicon cannot match. KoLo provides a candidate runtime layer for continuity, memory, orchestration, observability, and recovery around emerging biological-compute hardware.

Neuromorphic chip vendors

Edge-accelerated, spiking, low-power compute. KoLo’s substrate adapter interface accepts your inference endpoint. The daemon’s closed-loop scheduling matches the deterministic event-driven model your hardware was designed for.

HPC operators & research clusters

University supercomputers, national labs, and commercial HPC providers. KoLo runs as a long-lived service on your cluster, dispatches inference to your inference farm, and produces reproducible benchmarks for cross-substrate research.

Integration arc

From kick-the-tires to joint paper.

A phased engagement. Each phase has an explicit deliverable. Each phase is exit-able without lock-in. Joint experiments end in shared telemetry and a co-authored paper unless the partner elects otherwise.

  • Phase 0 · PilotKoLo Gateway + a single daemon agent on a partner-supplied substrate or inference endpoint. Continuity benchmark run. Telemetry shared.
  • Phase 1 · AdapterSubstrate adapter implemented against the partner’s hardware. Closed-loop scheduling validated. Recoverable state confirmed. Substrate transition test from silicon to partner-substrate.
  • Phase 2 · MeshMulti-agent peer mesh across partner substrate and silicon. Federated identity across lab nodes. Long-horizon autonomy run with telemetry.
  • Phase 3 · PaperJoint methodology paper. Reproducible benchmarks, substrate-adapter spec contribution, ethics-and-governance appendix. Co-authored, peer-review-targeted.

Bring the substrate.

We bring the runtime.

KoLo is engineered, tested, and waiting for the first lab that puts a non-silicon substrate underneath it. The engagement is phased, scoped per partner, and ends in a joint paper.

Discuss a lab pilot HPC & BioCompute Lab Read the paper