Episodic-to-Semantic Consolidation That Preserves Identity

Xue Qin, Simin Luan, Cong Yang and Zhijun Li (submitted 2 July 2026) propose treating memory consolidation as a deterministic transformation from episodic memory to a separately addressable semantic layer, f: M^ep -> M^sem, that updates knowledge without changing an agent's certified identity.

What do the authors propose?

The paper defines consolidation as a deterministic function f: M^ep -> M^sem that writes a semantic knowledge layer separate from the agent's manifest, and it proves identity invariance by showing the identity hash does not read M^sem. The authors present a formal account of the agent representation and a structural lemma on the manifest's hash-input set that secures the agent's certified identity against consolidation-driven mutation. They specify a deterministic aggregation algorithm whose outputs are auditable database rows containing explicit confidence values and supporting-event provenance.

How did they evaluate the construction?

The authors validated the design with synthetic experiments that checked per-field correctness and byte-equal identity across consolidation passes. Against a calibrated Bayesian-shrunk baseline, the construction produced a mean 79.82% reduction in unproductive planner attempts, with a 95% BCa confidence interval [78.02%, 81.49%] computed across 10 seeds. The running case study in the paper is an embodied service agent, illustrating how lessons accumulate as queryable facts in M^sem while the manifest and its certified identity remain unchanged.

Why it matters

The proposal separates knowledge accumulation from the signed identity manifest, resolving a practical tension for regulated autonomic agents that operate under audit and certification constraints. By making the semantic layer separately addressable and auditable, the method preserves cryptographic identity guarantees while allowing an agent to incorporate operational lessons. The reported 79.82% mean reduction in unproductive planner attempts implies fewer wasted planning cycles once consolidated knowledge is available as queryable facts.

What to watch

Adoption hinges on real-world tests beyond the paper's synthetic experiments: look for published evaluations on deployed embodied agents or third-party audits that reproduce the byte-equal identity property and the reduction in unproductive planner attempts. Another key signal will be whether the deterministic aggregation algorithm and its provenance schema are integrated into existing audited-agent workflows.

The paper is available on arXiv as "Episodic-to-Semantic Consolidation Without Identity Drift" (submitted 2 July 2026) by Xue Qin, Simin Luan, Cong Yang and Zhijun Li. It frames consolidation as a knowledge-update discipline in which lessons accumulate as auditable, queryable facts while the agent's certified identity remains byte-equal across its operational lifetime.