beyond retrieval: LeAP for coherent long-term agentic memory
Treating agent memory as epistemology, not retrieval.
An agent reads on Monday that a client prefers email for all communication. On Tuesday, it sees a newer note saying the same client now prefers Slack. Both get stored. Both get retrieved. Neither invalidates the other.
Weeks later, the agent confidently drafts an email because that fact ranked higher in similarity. The contradicting note sits in the same store, at the same confidence, saying the opposite.
This is not a retrieval problem. Both notes retrieved correctly. It is an epistemology problem: the system has no concept of contradiction, no mechanism for supersession, no way to propagate a correction.
RAG was built for static corpora. Long-running agents violate every assumption. They learn over time. They encounter contradictions. They build beliefs on beliefs. Hallucinations from session one resurface in session twelve as ground truth.
What if we treated memory as epistemology instead of retrieval?
Not “store and retrieve” but “what do I believe, why, and what would change my mind?” Stratified belief types. Evidence requirements. Provenance chains. Coherence enforced at write time, not hoped for at read time.
On 500 agent sessions: 95% contradiction detection vs 66% baseline. 87% revision propagation vs 12%. Corrections that actually cascade.
The paper formalizes this as Epistemic Augmented Generation and presents CITE, a four-layer architecture that implements it.