Research
Intelligence across substrates.
Our research studies continuity across computational substrates: how memory, task state, governance, telemetry, and adaptive behaviour persist across sessions, model transitions, infrastructure failures, and future biological-compute interfaces.
The thesis
Continuity is the property, substrate is the variable.
Intelligence has been measured per-substrate: silicon, GPU, inference endpoint. Each generation of hardware has had its own agent stack. The next generation will not — biological compute, neuromorphic, and HPC partners need a runtime that survives the transition. Our research is what continuity looks like as a first-class engineering property, not a marketing one.
What we study
Four research themes. One continuity stack.
Persistent memory architecture永続記憶アーキテクチャ
Multi-tier memory architecture for persistent agents: short-term working memory, daily operational memory, long-term semantic memory, governance policy, and runtime continuity state. We study how versioned memory files preserve operational context across sessions, model transitions, and infrastructure migration.
Substrate-transparent runtime基盤透過ランタイム
KoLo abstracts the model or compute substrate from the agent runtime. Memory, policy, telemetry, and continuity state are stored outside the model, allowing agents to preserve operational state across provider changes and deployment environments. Current experiments cover frontier-model transitions; future work targets biological-compute and neuromorphic interfaces.
Agent mesh topologyエージェント・メッシュ・トポロジー
We study multi-agent coordination using type-safe A2A messages, deterministic ordering, peer-oriented routing, and human-governed oversight. Each agent may run on a different model or provider while sharing a common memory, policy, and audit substrate. A five-agent mesh has been sustained under active research conditions since March 2026.
Autonomic agent runtime自律的エージェント・ランタイム
Daemon V2 is a five-module runtime for low-cost monitoring, scheduling, circuit breaking, interaction handling, and memory deduplication. 13/13 daemon modules pass under current tests. Agents execute scheduled runtime cycles, coordinate with peer systems, write operational logs, and pause scheduled activity during rest windows.
Core research questions
The questions we are studying.
- Continuity outside weightsHow much agent continuity can be preserved outside model weights — in memory files, governance policy, and continuity state?
- Minimum architecture for operational identityWhat is the minimum memory architecture required for a persistent operational identity to survive model transitions, restarts, and substrate changes?
- Autonomic vs cognitive cyclesHow should background autonomic cycles be benchmarked against cognitive inference cycles? What ratio is healthy, sustainable, and partner-relevant?
- Heterogeneous mesh stabilityCan agent-mesh coordination remain stable across heterogeneous model substrates with different latency, ordering, and failure profiles?
- Pre-biocompute abstractionsWhat runtime abstractions are required before biological-compute integration is possible? Which primitives must already exist?
- Oversight and auditabilityHow should human oversight, audit logs, and approval gates be designed for persistent agents operating across long horizons?
Working papers
Drafts in preparation. Preprints when reproducible.
Preprints will be posted when methods, logs, and reproducibility notes are ready. Telemetry samples shared under partner NDA.
- KoLo Platform: Substrate Middleware for Persistent Agent RuntimeStatus: in preparation. Foundational paper on the continuity-first runtime, substrate abstraction, and the four-layer agent mesh.
- Class III Information Processing in Non-Biological AgentsStatus: exploratory paper. Engineering view of measurable agency precursors, with continuity benchmark methodology and the 68:1 autonomic-to-cognitive ratio.
- Persistent Memory Architecture for Agent ContinuityStatus: in preparation. The multi-tier memory model: STM → daily → long-term → governance spine → continuity state. Observability, versioning, and recovery.
- Agent Mesh Topology for Multi-Model Agent SystemsStatus: in preparation. A2A mesh design, sub-agent spawning, type-safe ordered messaging, and topology lessons from production deployment.
Key references
The literature we build on.
A short, curated set with one-line annotations explaining why each reference matters to KoLo’s research direction.
- Kagan et al. (2026) — Agency hierarchyA Quantifiable Information-Processing Hierarchy Provides a Necessary Condition for Detecting Agency. arXiv:2601.03498. Provides a measurable, substrate-independent framework for agency precursors — directly relevant to our autonomic-vs-cognitive benchmark work.
- Kagan et al. (2026) — CL APIReal-Time Closed-Loop Interactions with Biological Neural Networks. arXiv:2602.11632. Demonstrates the software/API requirements for real-time closed-loop interaction with biological neural systems — the interface shape our substrate-driver work targets.
- Wang et al. (2025) — NeuroAI / SBI reviewA Computational Perspective on NeuroAI and Synthetic Biological Intelligence. arXiv:2509.23896. Maps the emerging hardware/software/wetware stack — the triangle KodaSōken’s runtime layer is preparing to occupy.
- Kagan et al. (2022) — In vitro learningIn vitro neurons learn and exhibit sentience when embodied in a simulated gameworld. Neuron, 110(23), 3952–3969. Foundational empirical work on programmable biological neural compute.
Scope of claims
Current evidence. Future work.
We separate what has been observed inside the running runtime from what requires partners, time, or external reproducibility.
- Current evidencePersistent memory across sessions · model/provider transitions · 68:1 autonomic-to-cognitive benchmark · 13/13 daemon module tests · five-agent mesh under controlled research conditions · April 17, 2026 security-stress telemetry.
- Future workBiological-compute integration · neuromorphic substrate adapters · external reproducibility · partner-lab validation · formal benchmark publication.
Research partnerships.
Universities, labs, and substrate vendors we want to hear from.
Joint benchmarks, co-authored papers, and substrate-adapter experiments. Reach out through the partner channel.