Does abstract preference knowledge outperform specific interaction recall?

The PRIME framework systematically compares episodic and semantic memory instantiations for LLM personalization, grounded in the cognitive dual-memory model (Tulving). The findings are consistent across model sizes and families:

Semantic memory > episodic memory. Using semantic memory (SM) alone — whether parametric (LoRA-encoded preferences) or textual (hierarchical summaries or parametric knowledge reification) — generally leads to higher personalization performance than using episodic memory (EM) alone. This suggests that abstract preference knowledge ("this user values concise factual responses") is more useful for personalization than retrieving specific past interactions ("the user asked about cats on Tuesday").

Recency > similarity for episodic recall. Within episodic memory, simple recency-based recall outperforms semantic-similarity retrieval in both accuracy and speed. The most recent interactions are the strongest predictors of immediate user behavior. This challenges the default design assumption that similarity-based retrieval is always superior.

Task fine-tuning > preference tuning. Among semantic memory instantiations, task-oriented fine-tuning (T-FT) — which directly learns the mapping from input query to desired outcome — achieves the best performance. Preference tuning methods (DPO, SIMPO) underperform, which deserves further investigation. Even input-only training (next token prediction, conditional input generation) achieves gains without task-specific labels, validating that semantic memory can encode useful preferences from raw user history alone.

Dual memory without mediation can backfire. Integrating both memory types without personalized thinking (DUAL) occasionally yields lower results than SM alone. This is a critical design warning: potential conflicts between episodic and semantic memories can be counterproductive if not properly mediated. Personalized thinking — synthesized reasoning traces that integrate both memory types — resolves this conflict and achieves superior performance.

The relationship to existing memory architectures is direct. Since How should agents decide what memories to keep?, the PRIME finding adds a hierarchy to that taxonomy: semantic memory should be the primary personalization signal, with episodic memory as a supplementary source that requires mediation to avoid conflicts. This inverts the common design pattern of treating episodic recall as the primary memory mechanism and abstracting only when retrieval is impractical.