Raw-RAG — raw-history retrieval as a memory baseline

TL;DR

Dense retrieval and reranking over raw conversation history outperform the compared agentic memory systems in 6 of 7 benchmark-model settings. This is evidence for retaining a simple baseline on the evaluated conversation tasks, not proof that all memory structures are unnecessary. (source)

Authors & Org

Zhengye Bao, Shengtao Zhang, Haokai Zhang, Shiqi Lou, Muning Wen and Minnan Luo. Affiliations are not established in the captured abstract record. First submitted 2026-10-22. (source)

Method

Raw-RAG retrieves raw historical passages densely, reranks them and supplies evidence to a reader LLM. The study compares this with memory systems using mutation, structural organization or learned retrieval, then examines evidence coverage and utilization. Useful evidence includes answer values and the relationships and dependency paths needed to derive answers. (source)

Results

Across LoCoMo, LongMemEval, MEME and MemoryAgentBench, the authors report leads of 0.13-5.53 points over the strongest compared agentic memory system. Higher coverage of necessary evidence can offset lower utilization. Even giving complete access to transformed memory stores leaves an accuracy gap against raw evidence. (source)

Significance

For Agents (LLM Agents) and Context Compaction, the practical implication is to compare memory transformations against access to preserved raw information. The result concerns long-term conversation understanding; it does not establish the best design for every autonomous workflow. (source)

Open Questions

The captured abstract does not specify the full model configurations, token budgets, retrieval costs or latency. Those details and the complete paper need review before attributing the gap to memory organization alone. No independent replication was performed. (source)

Cite

• arXiv:2610.30755 — first submission 2026-10-22.
• Captured abstract.