Is Complex Agentic Memory Always Needed? Why Naive RAG Beats Agentic Memory Systems

Last Updated: 24 October 2026

Key takeaways

• A plain retrieval baseline over raw chat history beat 12 agentic memory systems in six of seven benchmark-model settings across LoCoMo, LongMemEval, MEME and MemoryAgentBench.
• The margin over the strongest memory system ranged from 0.13 to 5.53 points, while ingestion of a one-million-token history took 0.33 hours instead of 3.25 to 26.95 hours.
• Raw history preserves both the answer value and the relations and dependency paths needed to derive it. Extracting facts or summaries tends to drop those links.
• Even complete access to a transformed memory store left systems below raw evidence, so better retrieval cannot repair information lost during ingestion.
• Adding atomic facts, summaries or graph expansion to the same raw baseline reduced accuracy by 5.78 to 19.62 points.
• Run the raw RAG baseline on your own logs before buying a memory layer, and add structure only where a measured coverage gap persists.

Agentic memory has become a default line item in AI architecture. Mem0, MemOS, EverMemOS, Zep and a dozen cousins promise to turn long chat histories into clean facts, graphs and summaries that an agent can recall on demand. The pitch is intuitive: raw transcripts are messy, so rewrite them into structured memory.

A paper published on arXiv on 22 October 2026 tests that assumption directly. According to Zhengye Bao and co-authors in "Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought," a naive dense-retrieval baseline over raw history ranks first among the compared memory systems in six of seven benchmark-model settings. The result is not a small edge. On MemoryAgentBench it leads the strongest memory system by 5.53 points, and it ingests a one-million-token conversation in 0.33 hours instead of 26.95 hours for the slowest system tested.

What did the Raw-RAG study actually test?

The study tested whether agentic memory operations add accuracy beyond simple retrieval over the original conversation. It compared Raw-RAG, a dense retrieval plus reranking pipeline, against 12 agentic memory systems and a full-context control. The benchmarks covered four different memory tasks: LoCoMo with 1,540 non-adversarial questions, LongMemEval with 500 examples, MEME with 1,188 phase-specific questions, and MemoryAgentBench with 2,071 questions drawn from 14 sources. Each system was run with the same reader model and evaluation procedure, so the comparison isolates how evidence reaches the reader rather than how clever the reader is.

The baselines included EverMemOS, MemOS, EMem, SimpleMem, MemU, MemMachine, Mem0, Memora, Nemori, MemoryOS, A-MEM and RecMem. That is close to a full census of the commercial and research memory layer market as of late 2026. The finding is therefore not a straw-man comparison. It is a test of the category.

How does Raw-RAG work?

Raw-RAG serializes each conversation with message timestamps and image captions, splits it into 256-token windows with 32-token overlap, and embeds every window into a flat vector index. At query time it embeds the question, retrieves the 128 closest windows by cosine similarity, reranks those candidates with a cross-encoder, and sends the top 15 windows to a reader LLM. There is no extraction, summarization, graph construction, consolidation or mutation at write time. The original text stays intact and remains the evidence the reader sees.

That design is almost aggressively boring. It is also the point. The paper deliberately isolates evidence access from memory management, so any accuracy difference can be attributed to what the memory layer did to the source material before retrieval.

What were the results across the four benchmarks?

Raw-RAG ranked first in six of seven benchmark-model settings when full-context was excluded. It led LoCoMo by 0.65 to 0.95 points, LongMemEval by 0.13 points, GPT-4o-mini MEME by 2.24 points, and MemoryAgentBench by 4.48 to 5.53 points. The one loss was MEME with DeepSeek-V4-Flash, where MemMachine led Raw-RAG by 1.52 points. Full-context also beat Raw-RAG by 19.98 points in that same setting, which shows that the reader can use a very large context well when the benchmark fits inside it.

"Raw-RAG ranks first among the compared memory systems in six of seven benchmark-model settings," the authors write. The consistency matters more than any single score. A naive baseline that wins across four benchmarks and two reader models is not a fluke of one dataset.

Why does raw history retrieval beat structured memory?

The paper separates two things that memory systems usually conflate: coverage and utilization. Coverage is the share of questions where the retrieved context contains enough evidence to answer. Utilization is accuracy conditional on that evidence being present. Across MemoryAgentBench, coverage ranged from 19.07% to 45.29%, while utilization stayed in a narrow band from 80.00% to 89.71%. On MEME, coverage varied by 40.66 points and utilization by 21.04 points. The systems differ far more in whether they supply the right evidence than in whether the reader can use it once supplied.

Raw-RAG had the highest coverage on both benchmarks, 45.29% on MemoryAgentBench and 76.35% on MEME. Its utilization was not the best, 84.43% and 73.76% respectively, but the coverage advantage more than compensated. That is the core mechanism: evidence access is the bottleneck, not reader sophistication.

The second mechanism is path preservation. Answering a question often requires more than a value. It requires the order, the update history or the dependency chain that makes that value current. On MEME tracking and cascade questions, Raw-RAG preserved a complete supporting path in 62.34% of cases where it had already found the answer value. MemMachine reached 56.70%, Mem0 47.66%, and MemU only 9.62%. A system can retrieve the right fact and still fail because it lost the relation that explains how the fact came to be.

What happens when a memory system gets complete access to its own store?

The study ran a complete-store oracle on MEME. Instead of relying on each system's native retrieval, it gave the reader gold-informed access to every phase-visible memory the system had exported, under the same 5,000-token packing budget. Every system improved, but the recovered accuracy depended on what the system had stored. Raw-RAG reached 57.54% and MemMachine 57.46%, compared with 61.60% for the raw-evidence control. Systems that had transformed history into facts or summaries recovered less.

When the packing ceiling was removed, all seven stores remained 7.58 to 29.38 points below raw evidence, and every paired 95% confidence interval excluded zero. The authors are careful to say this is not a pure measure of information loss, because evidence assembly and position effects such as lost-in-the-middle can also contribute. But the direction is clear: information discarded or distorted during ingestion is difficult to recover through retrieval alone.

How expensive is agentic memory compared with raw retrieval?

The accuracy result is only half the story. The cost difference is larger. On BEAM-1M, a single conversation approaching one million tokens streamed into each system with one external worker, Raw-RAG finished ingestion in 0.33 hours. MemOS, the next-fastest completed run, took 3.25 hours. Memora took 26.95 hours. That makes Raw-RAG roughly 81 times faster than the slowest system and about 10 times faster than the next-best system.

On a smaller MEME slice of 10 episodes and 110 questions, Raw-RAG completed add-to-ready plus search in 305.91 seconds. EMem took 1,561.97 seconds, Mem0 4,142.95 seconds, and EverMemOS 9,301.48 seconds. Raw-RAG was 5.1 times faster than the next system and 30.4 times faster than EverMemOS. The ablation add-ons were worse still: atomic facts, summary augmentation and graph expansion required 21.6 to 109.3 million LLM tokens and 2.9 to 8.8 hours, against 12 minutes for Raw-RAG.

Do memory-management add-ons improve the baseline?

No. The study added three common memory operations to the same Raw-RAG pipeline on MemoryAgentBench, reusing the chunking, embedding model and reranker. Atomic-fact extraction reduced test-time learning from 47.68% to 15.37% and total accuracy by 19.62 points. Summary augmentation lost 5.81 points, mainly on conflict resolution. Graph-based retrieval expansion lost 5.78 points. Even display-turn deduplication, a much lighter cleanup step, lost 2.39 points.

The ablations also show which parts of Raw-RAG actually matter. Removing the cross-encoder reranker cost 3.92 points. Returning only one retrieved item cost 11.73 points, and returning 120 items cost 14.57 points through noise. Final k values from 5 to 60 stayed within 1.3 points of the default of 15, and chunk sizes from 128 to 1,024 tokens made no significant difference. The lesson is that reranking and moderate retrieval depth carry the performance, not the exact chunk size or the size of the memory taxonomy.

When should you still use an agentic memory layer?

Agentic memory is not useless. The paper's scope is offline conversation understanding on public benchmarks. It does not test interactive, multimodal or broader operational settings, and the authors explicitly list storage and privacy as open concerns. Raw history is expensive to store and harder to delete selectively, which matters under privacy regimes and customer deletion requests.

A structured layer still makes sense when raw retrieval cannot fit the context budget, when retention rules require deleting specific facts rather than whole conversations, when the agent must share memory across sessions with different access scopes, or when a weaker reader needs explicit state maintenance. The paper's own future work asks whether structured memory helps weaker readers or reduces repeated reasoning across queries. Those are the conditions to test, not assumptions to bake in.

How do you run the Raw-RAG baseline on your own data?

You can reproduce the baseline in an afternoon with an existing vector store and a reranker. The important discipline is to evaluate evidence coverage separately from answer accuracy, because that is where the diagnostic signal lives. If your current memory layer cannot beat raw retrieval on coverage, the added complexity is not earning its keep.

1. Export raw conversations with timestamps, speaker roles and any image captions. Keep the original text, not a summary.
2. Split each conversation into 256-token windows with 32-token overlap and embed every window into a flat vector index.
3. At query time, retrieve the 128 closest windows by cosine similarity, then rerank with a cross-encoder and pass the top 15 to your reader model.
4. Score answers on a held-out question set, and separately label whether the retrieved context contained sufficient evidence. Report coverage and utilization as two numbers.
5. Run the same questions through your current memory system. Compare coverage first, then accuracy, latency and ingestion cost. Add structure only where coverage is materially worse.

What does this mean for Australian teams building AI agents?

For most Australian teams building internal assistants, the practical takeaway is to stop treating a memory layer as a mandatory component. A raw-history retrieval baseline is cheaper to build, cheaper to run and easier to audit. It also keeps the source evidence intact, which matters when a customer asks why the agent said something and you need to show the exact conversation rather than a derived fact.

In our own OpenClaw deployment, the agent keeps a curated memory file for durable preferences and operating rules, but falls back to raw session-history search when a detail is missing. The curated layer is fast and compact. The raw layer is what actually answers questions like "what did we decide on 12 October" or "which client asked for that change." That matches the paper's finding: the summary is useful for orientation, but the raw evidence is what resolves the question.

The cost case is strong for local deployments too. If you are running models on your own hardware, ingestion time is your own compute. A memory pipeline that spends 3 to 27 hours of LLM calls rewriting a million-token history is a real operating cost, not a rounding error. Raw chunking and embedding is minutes of work by comparison.

Frequently asked questions

Is RAG better than agentic memory for long-term conversations?

In the October 2026 benchmark study, yes. Raw-RAG, a dense retrieval plus reranking baseline over raw history, ranked first in six of seven benchmark-model settings and beat the strongest agentic memory system by 0.13 to 5.53 points. The exception was MEME with DeepSeek-V4-Flash, where MemMachine led by 1.52 points.

Do I need a memory layer like Mem0 or MemOS for my AI agent?

Not by default. Run a raw-history RAG baseline on your own conversations first. If it matches or beats your current memory system on evidence coverage and answer accuracy, you do not need the extra write pipeline. Add a memory layer only where raw retrieval cannot meet context, privacy or access-control requirements.

How much does an agentic memory system cost to run?

In the study, ingesting a one-million-token history took 0.33 hours with Raw-RAG, 3.25 hours with MemOS and 26.95 hours with Memora. On a 110-question MEME slice, add plus search took 306 seconds for Raw-RAG versus 1,562 to 9,301 seconds for the memory systems. The memory-management ablations also consumed 21.6 to 109.3 million LLM tokens.

Why does raw history retrieval work so well?

Raw history keeps both the answer value and the relations and dependency paths needed to derive it. The study found that memory systems differ far more in evidence coverage than in reader utilization. Raw-RAG achieved 45.29% coverage on MemoryAgentBench and 76.35% on MEME, the highest of the compared systems, and preserved a complete supporting path in 62.34% of cases where it found the answer value.

When is agentic memory still worth using?

Agentic memory is worth testing when raw history exceeds your context or storage budget, when deletion and retention rules require selective removal of facts, when different sessions need different access scopes, or when a weaker reader needs explicit state maintenance. The paper's evidence covers offline conversation benchmarks, not interactive or multimodal production systems.