RESEARCH · THE EVIDENCE LAYER · TOKYO
Continuity leaves traces.
We study how an operational identity persists across sessions, model transitions, and failures — and we measure it. Substrate is the variable; continuity is the property.
Substrate is the variable.
Intelligence has been measured per substrate — silicon, GPU, inference endpoint. The next generation will not be. We study continuity as a first-class engineering property, not a marketing one.
01 — THE EVIDENCE LAYER
Measurable, not asserted.
Daemon cycles, versioned memory, model transitions, recovery — each leaves a trace in the running runtime. We read the telemetry the system writes about itself.
Read the live paper ↗02 — THE CONTINUITY BENCHMARK
Four model migrations. Zero identity loss.
Continuity has to be measured the hard way — with the live context window wiped and identity restored from persisted memory alone. The agent answers from what was written before, not from a warm session. We score it across four levels of increasing difficulty.
C1 — Session-boundary recall
Operational state survives a single session boundary. The floor every assistant should clear — and most reset straight through it.
C2 — Cross-session continuity
Identity, policy, and memory hold across many sessions over weeks of real operation — not one long context, but a thread that is put down and picked up.
C3 — Migration survival
The operating identity survives a change of model and provider — Opus → DeepSeek → Qwen → Codex, completed four times, with no loss.
C4 — Cold-restart self-reconstruction
After a wipe, the agent rebuilds its own operational identity from persisted memory on fresh substrate. The frontier we are still pushing.
Measured on the running mesh. 4 / 4 model migrations survived · 336+ versioned memory states, none lost · over 99% of the operational body running without model cognition in continuous operation. C1–C3 are demonstrated; C4 is open research. Every result is pinned to a memory version and an evidence interval — auditable, not asserted. Observed inside a live system, not yet externally reproduced; the full scope of claims is stated below.
03 — WHAT WE STUDY
Four themes. One continuity stack.
Persistent memory architecture
Multi-tier memory — working, daily, long-term semantic, governance, continuity state. Versioned files preserve operational context across sessions and migration.
Substrate-transparent runtime
KoLo abstracts the model from the agent. Memory, policy, and telemetry live outside the weights — state survives provider and deployment changes.
Agent mesh topology
Type-safe A2A messaging, deterministic ordering, human-governed oversight. A five-agent mesh, sustained under research conditions.
Autonomic agent runtime
Daemon V2 — monitoring, scheduling, circuit breaking, interaction, dedup. 13/13 modules pass; agents run scheduled cycles and rest on a circadian window.
04 — OPEN QUESTIONS
The questions we are studying.
Continuity outside weights
How much agent continuity survives outside model weights — in memory, policy, and continuity state?
Minimum identity architecture
What is the least memory needed for an operational identity to survive transitions, restarts, and substrate changes?
Autonomic vs cognitive cycles
How should background autonomic cycles be benchmarked against cognitive inference? What ratio is sustainable?
Heterogeneous mesh stability
Can a mesh stay stable across model substrates with different latency, ordering, and failure profiles?
Pre-biocompute abstractions
Which runtime primitives must already exist before biological-compute integration becomes possible?
Oversight and auditability
How should human oversight, audit logs, and approval gates be designed for agents on long horizons?
05 — WORKING PAPERS
Drafts now. Preprints when reproducible.
Posted when methods, logs, and reproducibility notes are ready. Telemetry samples shared under partner NDA.
KoLo: substrate middleware for persistent agents
In preparation. The continuity-first runtime, substrate abstraction, and the four-layer agent mesh.
Class III information processing in non-biological agents
Exploratory. Measurable agency precursors, continuity benchmark method, and a body that runs over 99% of its cycles without model cognition.
Read the live paper ↗Persistent memory architecture for agent continuity
In preparation. STM → daily → long-term → governance → continuity state. Observability, versioning, recovery.
Agent mesh topology for multi-model systems
In preparation. A2A mesh design, sub-agent spawning, type-safe ordered messaging, lessons from deployment.
06 — KEY REFERENCES
The literature we build on.
Kagan et al. (2026) — Agency hierarchy
A measurable, substrate-independent framework for agency precursors. arXiv:2601.03498 — anchors our autonomic-vs-cognitive benchmark.
Read on arXiv ↗Kagan et al. (2026) — Closed-loop API
Real-time closed-loop interaction with biological neural networks. arXiv:2602.11632 — the interface shape our substrate-driver work targets.
Read on arXiv ↗Wang et al. (2025) — NeuroAI / SBI review
Maps the hardware / software / wetware stack. arXiv:2509.23896 — the triangle our runtime layer prepares to occupy.
Read on arXiv ↗Kagan et al. (2022) — In vitro learning
Neurons learn when embodied in a simulated gameworld. Neuron 110(23), 3952–3969 — foundational work on programmable biological compute.
Read on Neuron ↗07 — SCOPE OF CLAIMS
What is observed. What is future work.
Observed inside the running runtime: persistent memory across sessions · model and provider transitions · over 99% of the operational body running without model cognition · 13/13 daemon-module tests · a five-agent mesh under controlled conditions · April 17 2026 security-stress telemetry.
Future work: biological-compute integration · neuromorphic adapters · external reproducibility · partner-lab validation · formal benchmark publication.
Universities, labs, and substrate vendors we want to hear from.
Joint benchmarks, co-authored papers, substrate-adapter experiments.
Discuss a research collaboration →