New Method Consolidates Agent Knowledge Without Identity Drift

Long-running adaptive intelligent agents face a fundamental challenge: how to consolidate new knowledge without changing their core identity, especially in regulated environments where an agent's identity might be cryptographically certified. Traditional memory consolidation often involves fine-tuning the model or rewriting prompts, which effectively changes the agent. This research introduces a new paradigm where knowledge consolidation is treated not as a mutation of the agent's planner or identity manifest, but as a deterministic function that transforms episodic memory into a separately addressable semantic knowledge layer. Crucially, the agent's identity hash does not incorporate this semantic layer, ensuring that knowledge updates do not alter its certified identity. The paper provides a formal account of this agent representation, proving identity invariance through a structural lemma. It specifies a deterministic aggregation algorithm that produces auditable database rows with explicit confidence and provenance. Synthetic experiments validate the approach, demonstrating byte-equal identity across consolidation passes and a significant reduction in unproductive planner attempts, making it suitable for autonomic agents operating under strict audit contracts.