Ilmu Komputer & AI editorial
Runtime-Independent Persistent Agents: Preserving Identity, Memory, and Code Across Models, Harnesses, and Servers
The core problem
Long-lived AI agents are increasingly expected to operate over extended periods, yet contemporary agent stacks tightly couple identity, memory, and executable code to specific models, orchestration harnesses, interaction surfaces, and hosts. When any of these components is replaced—whether due to model deprecation, harness upgrades, or infrastructure migration—the agent's continuity is typically lost. Zhao and Zhao address this gap by asking whether an agent can retain its identity, memory, and executable body lineage while its runtime components are swapped out.
The paper introduces a runtime-independent architecture that explicitly separates a continuity-bearing substrate from replaceable execution and interaction bindings. Formally, the substrate is defined as , where denotes identity, denotes memory, and denotes the executable body lineage at time . The architecture is governed by six continuity invariants and a six-phase protocol: quiesce, checkpoint, validate, bind, rehydrate, and resume. The authors implement this design in a system called Enoch, which provides a reusable software body, private installed state, provider contracts,
Innovation
The operational case demonstrates that an established instance retains its identity, nonempty memory, body revision, and historical task IDs across host, joint Codex-to-Claude model/harness, and chat-surface substitutions. This provides direct evidence that the continuity substrate survives simultaneous changes to multiple runtime components. A separate cross-host task successfully resumes from a verified artifact checkpoint, showing that the checkpoint–rehydrate mechanism works across different hosts.
In the supervised Codex–Muse–Codex study, the system completes five ordinary round trips and five matched same-runtime controls on their first attempts after a separate qualification run. This indicates that the migration protocol does not introduce first-attempt failures in the tested workflow. The planned worker interruption test recovers through the native task API with the checkpoint preserved, a stale reply rejected, and one task completion recorded. This demonstrates that the system can handle mid-execution interruptions without losing continuity or duplicating effects in the tested scenario. Across all studies, the results support mechanical continuity for the tested deployme
Why it matters
The paper's central contribution is the architectural separation of continuity from runtime, which reframes agent persistence as a substrate-level concern rather than a property of any particular model or harness. The six invariants and the quiesce–checkpoint–validate–bind–rehydrate–resume protocol provide a principled foundation for reasoning about state preservation, capability changes, and continuation authority. Enoch's implementation demonstrates that this separation is feasible in practice, at least for the bounded workflows tested.
However, the authors are careful to bound their claims. The evidence supports mechanical continuity—the preservation of identity, memory, and body lineage—but not behavioral equivalence, meaning the agent may behave differently after migration even if its substrate is preserved. Arbitrary-task portability is not established, as the tested workflows are limited in scope. General exactly-once external effects remain an open problem, since the system cannot guarantee that external side effects are neither duplicated nor lost across migrations in all cases. Unattended reliability is also not demonstrated, as the live studies involved supervision. These limitations point to important directions for future work, including broader task coverage, stronger guarantees for external effects, and long-duration unattended operation. The architecture nonetheless provides a valuable foundation for building persistent agents that can evolve with their runtime environments.
Who should read this
Opening member content…