The agent that remembers you becomes an agent that drifts from itself. Separate what resets from what persists: let identity load fresh each session while knowledge survives in explicit storage. Predictability and relationship need not be traded — they live in different layers.
An agent that accumulates identity across sessions absorbs user patterns, drifts from its designed baseline, and creates security exposure through persistent credentials. An agent that resets completely each session cannot build relationships or learn preferences. Neither pole serves the designer who needs both.
An AI agent serves the same users repeatedly over time — a personal assistant, a domain expert, a customer service agent with ongoing relationships. The designer must choose whether the agent’s behavioral identity resets with each session (ephemeral) or accumulates across sessions (persistent). The choice has security implications, behavioral stability implications, and relationship-building implications that pull in different directions.
Separate what resets from what persists. Identity resets each session — the behavioral baseline, persona framing, and operational constraints load fresh from version-controlled configuration. Knowledge persists explicitly — user profiles, learned preferences, task history, and relevant context are stored in external semantic memory and retrieved at session start.
This is the “stateful serverless” architecture: the execution environment is ephemeral, but state management is explicit and external. Agents retain memory and contextual understanding between sessions without the economic and security costs of persistent execution.
The reset/persist boundary is the key design decision. Explicit storage decisions (what gets written to semantic memory and why) keep the agent from absorbing implicit behavioral drift while enabling legitimate adaptation.
Credentials and permissions are always session-bound. Cryptographic attestation binding agent identity to runtime environment and execution claims, with credentials automatically expiring on mission completion, prevents the credential accumulation vulnerability.
Ephemeral identity with explicit knowledge persistence requires infrastructure for semantic memory management. The agent cannot implicitly learn — every form of persistence must be designed as an explicit write. This is higher implementation cost but also higher transparency: the system always knows what persists and why.
The behavioral baseline remains testable and verifiable. Any session can be evaluated against the same specification.
Relationship continuity comes from the knowledge layer, not the identity layer. Users experience consistency through remembered context, not through a “self” that accumulates character across sessions. For some use cases this distinction matters to users; for others it does not.
Security is cleaner: a session-bound identity with expiring credentials cannot be exploited by delayed trigger attacks. The attack surface is bounded by the session.
The “Invitation Is All You Need” Black Hat attack demonstrated the risk of persistent agent context: calendar poisoning placed malicious commands that executed days later when the persistent agent encountered the triggering event. Session-bound identity would have prevented the delayed execution by expiring context with the original session.
Production agent memory systems (Letta/MemGPT, NVIDIA ICMS) now treat memory as a first-class, explicit architectural component rather than an implicit accumulation. Letta positions memory as “a first-class, explicit component of agent state.” NVIDIA’s Inference Context Memory Storage proposes a dedicated tier for “ephemeral but latency-sensitive” KV cache distinct from “durable and cold” compliance data.
The garden’s commission architecture uses ephemeral identity by design: each Gardener session reloads from the agent file, ensuring no cross-commission contamination. Learning persists explicitly through /deep-learning and session-log, not through model state accumulation. This is the stateful serverless pattern in practice.
extracted_from::[[Persona and Agent Personalities]]↑