Across agent memory and RAG, the dominant finding has inverted from ‘retrieve more’ to ‘retrieve selectively’: indiscriminate context injection degrades performance, and how retrieved material is structured and governed matters more than raw retrieval accuracy.
Evidence
- Memory-dosage and MESA results show selective, task-adaptive retrieval matches or beats full guideline injection while cutting tokens (e.g., 8.5% accuracy gain with 41% fewer tokens), and excessive retrieval harms sequential decision-making
- Beyond-Retrieval trajectory reuse and CABLE show post-retrieval structuring/linking outweighs retrieval quality alone
- Semantic/vector retrieval underperforms structured or symbolic memory on certain queries (Ontology-Grounded 0.98–1.00 vs 6–27%; MemoryLake beats vector RAG on math/physics)
- Governance failures dominate long-horizon reliability: compaction silently drops session constraints ~83% of the time, and unmanaged records that contradict or become obsolete corrupt downstream reasoning
Implications
- Investment should shift from bigger retrieval/memory stores toward selection policies, structured substrates, and constraint-preserving compaction
- Serving cost and reliability now hinge on internal memory behavior, so per-tier calibration and fail-closed governance become core design requirements rather than add-ons
Concepts
Agent Memory RAG Context Engineering Embeddings Agents
Confidence
high