Episodic-to-Semantic Consolidation Without Identity Drift

Long-running autonomous agents face a fundamental stability-plasticity dilemma: they must adapt to lived experience to improve performance, yet they are often bound by strict regulatory and audit contracts that require a stable, cryptographically certified identity. Conventional consolidation methods—such as fine-tuning, prompt rewriting, or policy distillation—mutate the agent's core artifacts, effectively forcing a re-certification event every time the agent learns. This paper asks: can an agent consolidate operational knowledge from episodic experience while keeping its identity certificate byte-equal?

The authors introduce a two-store memory architecture that separates the identity manifest from operational memory. The agent's identity is defined by a fixed-length cryptographic hash of a manifest. The authors prove, via a structural lemma, that if the consolidation function writes only to a separate semantic memory layer—and this layer is excluded from the manifest's hash-input set—the agent's identity remains invariant across any number of consolidation passes. They implement a deterministic, idempotent aggregation algorithm that converts raw episodic event logs into structured semantic facts (e.g., success rates, confidence scores, and provenance pointers) that downstream planners can use for grounding.

The framework successfully decouples knowledge updates from identity maintenance. Empirical validation on a synthetic benchmark demonstrates that this approach reduces unproductive planner attempts by a mean of 79.82% (95% BCa CI [78.02%, 81.49%]) compared to a non-memory baseline. Crucially, the agent's identity hash remains byte-equal across all consolidation passes, confirming that the agent can become operationally smarter without triggering a re-certification event. The system is computationally efficient, with consolidation latency scaling sub-linearly and semantic query times remaining in the sub-millisecond range.