New Architecture Lets AI Agents Keep Their Identity and Memory Across Different Models and Servers

Researchers have introduced an architectural approach that enables AI agents to persist their core identity and memories even when migrating to different underlying models, hosting systems, or interaction platforms. Rather than tying an agent to a single runtime environment, the new design separates the agent's essential state from the replaceable technical infrastructure around it.

The Core Architecture#

The system treats each agent as having a continuity-bearing substrate made up of three elements: a unique identity, accumulated memory, and the agent's executable code body. This persistent core remains stable while the external bindings—the orchestration framework, the language model powering it, and the interface through which users interact—can be swapped out without breaking the agent's continuity.

The approach is governed by six invariants that ensure state is preserved correctly and outlines a formal sequence for transitioning between environments: quiesce, checkpoint the current state, validate integrity, bind to the new runtime, rehydrate the agent's internal state, and resume operation.

Implementation and Testing#

A system called Enoch puts this design into practice using reusable code bodies, isolated installed state, standardized provider contracts, and secure migration procedures. The researchers tested the architecture across multiple scenarios:

• Identity preservation: An operational deployment maintained agent identity, memory contents, code revisions, and task history while switching from one hosting server to another and replacing the underlying models (moving from Codex to Claude) alongside a new chat interface.

• Cross-system recovery: A separate test confirmed that agents could resume execution from a validated checkpoint on a different host.

• Model switching: In a controlled study, agents completed multiple round-trip migrations between different model providers, and matched control runs succeeded on first attempt after an initial qualification phase. A simulated worker interruption demonstrated recovery through the task API, with stale responses appropriately rejected and task states correctly recorded.

Current Scope and Limitations#

The evidence gathered supports mechanical continuity within the tested deployments and bounded workflows. However, the researchers acknowledge significant open questions: the work does not yet demonstrate behavioral equivalence across different model switches, establish that arbitrary tasks remain portable, guarantee exactly-once semantics for external effects, or prove the system's reliability for unattended long-term operation. These remain challenges for future work as the approach matures.