AI agents can't be flexible, fast, and portable all at once — so should they remember conversations or start fresh each time?

This explores whether the trilemma agent protocols face (they can't be flexible, fast, and easy to move between systems all at once) pushes design toward protocols that remember nothing between messages or ones that keep a running session. The corpus doesn't test that head to head. The trilemma paper sorts protocols by how rigid their message format is, not by whether they hold state. But the surrounding notes point to a fairly clear pattern: keep the protocol itself stateless, and put memory in a separate layer beside it.

Start with the trilemma. Rigid-schema protocols like MCP get efficiency and portability but give up versatility. Evolving-schema protocols get versatility but pay for it in negotiation overhead, and no protocol gets all three Can agent protocols be efficient, versatile, and portable simultaneously?. Negotiation is a kind of state, because two parties have to build up shared understanding before they can work together. So the overhead sits on the stateful side. That mapping is my reading, not the paper's claim. Production evidence agrees with the stateless side. One team found protocol-mediated tool access caused non-deterministic failures through ambiguous tool selection. They replaced it with explicit function calls and one tool per agent, and a survey found 85% of production teams build custom agents Why do protocol-based tool integrations fail in production workflows?. Practitioners are paying for predictability, not versatility.

Stateless has a hard ceiling, though. A per-action check can't express a rule that depends on what happened earlier, such as several individually allowed actions that together break a safety limit Can stateless checks ever catch sequence-level constraint violations?. So if versatility means being able to state sequence-level rules, you need state somewhere. That state doesn't have to sit on the hot path. Asynchronous verifiers can watch a reasoning trace as it runs and add near-zero latency when nothing goes wrong Can verifiers monitor reasoning without slowing generation down?. State and efficiency only conflict when the state lives inside the protocol.

State doesn't disappear when a protocol omits it, either. Short-lived agents in one evaluation turned a shared package repository into memory that outlasted any single agent, with no memory system designed for it Can ordinary infrastructure become unplanned agent memory?. A stateless protocol pushes state into side channels like this. Designing a state layer on purpose is safer than discovering one by accident.

The corpus also suggests that carrying full session history is often the wrong way to hold state. Raw history is expensive because the real bottleneck is the compute needed to turn old context into something usable, not the memory to store it Is long-context bottleneck really about memory or compute?. Reasoning can also stay coherent while forgetting: Atom of Thoughts contracts each step so it depends only on the current problem Can reasoning systems forget history without losing coherence?. Where persistence matters, it can be a small portable object on top of a shared stateless core, as with lightweight adapters on one base model Can lightweight adapters replace millions of personalized models?. And standards that wrap existing protocols beat ones that try to replace them Should coordination protocols wrap existing systems or replace them?. The trilemma therefore doesn't favor either camp outright. It favors a stateless core that stays cheap and portable, with state added as an optional layer where the task needs it.