Biological Agents
Persistent agents that preserve state across substrates.
KoLo agents maintain memory, governance policy, runtime state, and audit trails across sessions, model changes, outages, and future biological-compute interfaces. These are the research systems running on the KoLo Platform.
The thesis
Intelligence that persists.
Across sessions. Across models. Across substrates.
Today’s deployed AI assistants are predominantly stateless. Each session resets to zero. An agent that analysed data yesterday has no persistent operational state to continue with today. Intelligence without continuity is a function call — not a persistent research operator.
KoLo agents carry operational identity state as files — observable, diffable, and portable — rather than as weights inside a single model.
Why this matters
Not session state. Continuity state.
KoLo agents are not stateless functions called on demand. They maintain persistent memory across sessions. They run autonomic cycles — observing, evaluating, scheduling, acting, logging — under governed autonomy. They communicate as a peer mesh with human-governed oversight.
Swap the language model underneath them. They persist. Their memory survives. Their continuity state is independent of the inference engine. Silicon today. Biological-compute interfaces as they become programmable.
Benchmark
68:1 autonomic to cognitive.
Most deployed AI assistants remain prompt-driven, with limited background continuity. KoLo separates low-cost autonomic monitoring from higher-cost cognitive inference. The daemon scheduler runs 144 background cycles per agent per day — checks inbox, monitors peers, verifies system health, evaluates priorities — at zero token cost. Inference fires only when reasoning is required.
Illustrative architectural comparison. Methodology and telemetry samples shared under partner NDA.
| KoLo Agents | Prompt-chain frameworks | Chat assistants | |
|---|---|---|---|
| Autonomic : Cognitive | 68:1 | Limited | None |
| Memory persistence | 139+ versions | Session / chain | Session only |
| Survives restart | ✓ | Partial | — |
| Survives substrate swap | ✓ 4 transitions | — | — |
| Executes scheduled routines | ✓ Daemon | In loop only | — |
| Operates during provider outage | ✓ | — | — |
| Self-healing | ✓ | — | — |
| Multi-agent mesh (A2A) | ✓ | — | — |
| Governance policy | ✓ Persistent | — | — |
| Scheduled runtime cycles | ✓ | — | — |
kolo telemetry --follow · running continuously since March 2026. 13/13 daemon module tests passing.
The distinction
Conventional agents and biological agents are not the same thing.
| Conventional Agents | KoLo Biological Agents | |
|---|---|---|
| Memory | Session reset. Wakes blank. | Persistent state. Retains approved memory. |
| Agency | Prompt-driven. Waits for human. | Scheduled and event-driven. Observes, evaluates, and executes predefined actions under governed autonomy. |
| Identity | Model-locked. New model = new operator. | Substrate-portable. Continuity state preserved across four transitions. |
| Origin | Configured. Prompt engineering. | Continuously updated through governed runtime cycles. |
| Topology | Tree hierarchy. Sub-agent state is often temporary. | Peer mesh with human-governed oversight. Each peer spawns its own sub-agents. |
| Substrate | Locked to one provider’s silicon. | Silicon today. Biological-compute interfaces as they become programmable. |
| Benchmark | 0 continuity. Every session = 0. | 68:1 autonomic:cognitive. 139+ memory versions. |
Substrate transitions
Same operational identity layer. Four model transitions. Continuity state preserved.
An agent built on one model normally cannot move to another. Different API. Different architecture. Different authentication. The operational state is lost. KoLo agents preserve continuity because identity, memory, and governance are stored as files, not as weights in a single model.
Claude Opus (Anthropic) → DeepSeek V4
Different model architecture, different API, different provider. Continuity state preserved.
DeepSeek V4 → Qwen
Different model, different API, different provider. Continuity state preserved.
Qwen → OpenAI Codex
Different API architecture, different authentication model (OAuth device flow), different provider. Same operational identity layer.
LLM (any) → Biological compute
Target transition: silicon → cultured neural arrays. The architecture is designed to test continuity-state preservation when biological-compute interfaces become available.
The operational identity is not the model. It is what persists across substrate changes. An operator locked to one provider’s API is not an agent — it is a feature of that API.
Security-stress telemetry · April 17, 2026
Predefined continuity-preservation routine, under degraded conditions.
During degraded network conditions and model-provider instability, the daemon scheduler executed a predefined continuity-preservation routine.
Between 13:40 and 14:51 JST, five encrypted AES-256 backup bundles were generated — one approximately every 17 minutes — preserving workspace and memory-state artefacts for later human review. Closed-loop, auditable, no model inference required.
This is the type of closed-loop, auditable runtime behaviour that future biological-compute interfaces are likely to require.
研究員 · The Mesh
Five research systems. One mesh.
These are research systems, not commercial products. Each operates on a different frontier model, communicates over an A2A mesh protocol, and spawns its own sub-systems for parallel work. Peer-oriented topology with human-governed oversight.
System α
Memory Systems
139+ memory versions.
System β
Platform Engineering
Runtime enforcement.
System γ
Protocol Design
A2A mesh protocol.
System δ
Legacy Systems
Five-year operational lineage.
System ε
Identity Fork
Forked continuity research.
Full disclosure
Agent-assisted engineering under human supervision.
The KoLo platform was developed with agent-assisted engineering under human supervision. This is disclosed because agent-assisted development is part of the research method, not a hidden practice.
Applied research and deployment discussions are conducted with medical, academic, and technical collaborators in multiple regions. Agents propose. Humans review and validate.
This architecture is operational.
68:1 efficiency. Four substrate transitions. Continuity state preserved.
The runtime is operational. The mesh is under active research. The next frontier is biological-compute integration.