Reconciling Memory and Evidence: A Comparative Review of Latent Experience Retrieval and Prior–Context Arbitration in LLM Agents

Two recent lines of work address a common underlying problem from opposite ends: how should a large language model (LLM) reconcile what it already believes with new information it retrieves? Feng et al. (2026) tackle this at the level of an agent's own past experience, proposing ExpWeaver, a framework that encodes prior trajectories in latent space and fuses them into generation through learned cross-attention and a gated residual, trained end-to-end with reinforcement learning. Sun et al. (2026) tackle the same underlying tension from a diagnostic angle, introducing PAVE, a testbed that stratifies a verifier's pre-evidence epistemic state and measures how reliably it persists in correct beliefs or corrects wrong ones once retrieved evidence is introduced. This paper reviews both contributions in detail, extracts their shared conceptual core — the arbitration between parametric priors and external context — and argues that they are complementary rather than competing. We summarize each paper's methodology, experimental setup, and results at length, situate both works within the broader literature on retrieval-augmented generation, agent memory, and knowledge-conflict evaluation, and synthesize cross-cutting insights and open questions that neither paper addresses in isolation. We conclude with a discussion of practical implications for anyone building or evaluating RAG-based agents, and a set of concrete open research questions the two papers jointly motivate but neither answers. This is a review and synthesis article; it reports no new experiments, and all empirical claims are attributed to the original sources.