AI Agent Arxiv Digest — 2026-10-11

🌏 中文版

Today's Overview

Today's two papers happen to puncture the same assumption from two different angles: treating every memory the same way at retrieval time is no longer good enough. MemoType's answer is to classify first — memory isn't homogeneous; events, personal semantic facts, and general knowledge are three different kinds of things, and the system should identify the type before picking a retrieval strategy. CogMem's answer is to separate layers — facts and the attributed opinions people voice about them often collapse together in vector space, so the system should keep them apart in a graph structure, then let retrieval itself become a loop that decides its own next move instead of a one-shot similarity lookup. Both papers happened to be accepted the same week, to NeurIPS 2026 and EMNLP 2026 main conference respectively, and both use ablations to cleanly isolate what the "extra step of judgment" actually buys — solid evidence, though still worth watching for independent replication.

Terms Worth Knowing Before You Read

Paper One｜Splitting Memory by Type: Letting the Agent Decide How to Search

Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies Yi Wen, Derong Xu, Pengyue Jia et al. (City University of Hong Kong + Huawei Noah's Ark Lab) · arxiv: 2610.11573

Links: arxiv · alphaxiv

TL;DR

A learned router first classifies each memory and each query, then routes it to a matching retrieval strategy; the method beats nine baselines across four long-term memory benchmarks, is accepted at NeurIPS 2026, and leads the strongest baseline by 16.18 points of Recall@1 on LongMemEval-M.

Editorial Judgment

Field Context

RAG-style memory is the current mainstream approach for long-term dialogue agents, but most methods treat every memory the same way, applying one similarity or query-rewriting strategy to all of them. The authors note that on LongMemEval-S, episodic memories make up only 12.35% of instances while semantic memories dominate overwhelmingly — and a t-SNE visualization of the semantic cluster reveals a clear gap between personal and general knowledge. Memory, in other words, isn't homogeneous; using one strategy for different types of memory is like using the same key on different locks.

Mid-Level Walkthrough

• The problem: Imagine your agent needs to answer "what was the name of that restaurant I went to last time" versus "how long do cats typically live" — the first requires precise matching against your personal conversation history (episodic memory), the second is closer to retrieving general knowledge (semantic memory). Most systems use the same embedding similarity for both, like using the same move to answer two completely different kinds of questions.
• The method: MemoType first uses a learned router to classify each memory and each query into one of three types — episodic, personal semantic, or general semantic — then dispatches the matched pair to a pre-designed retrieval strategy (one of key expansion, hypothetical query rewriting, or hypothetical memory generation). A query-conditioned memory pruning module then filters out retrieved-but-irrelevant fragments. The authors also prove theoretically that using any single strategy on a multi-class memory corpus is subject to a fundamental upper bound on expected retrieval precision.
• Why it matters: This shows memory management can gain an extra "classify, then route" layer without retraining the underlying LLM — a meaningful boost from fixing the routing logic alone. That's especially appealing for teams already running an agent framework in production who don't want to touch the core architecture.

Key Findings

• Compared across 4 benchmarks (LongMemEval-S/M, LoCoMo, PerLTQA) × 9 baselines (HyDE, Mill, Query2Doc, SeCom, A-Mem, HippoRAG2, RAPTOR, MemoryOS, LightMem), MemoType achieves the best Recall@1 on all four datasets
• On LongMemEval-M, Recall@1 reaches 48.09%, 16.18 points above the next-best SeCom (31.91%)
• On PerLTQA, Recall@1 reaches 71.92%, 14.73 points above the second-best MemoryOS (57.19%)
• Ablation shows the type-aware strategy beats every single-strategy baseline, gaining 5.24, 5.14, and 5.68 points of Recall@1 respectively across the three datasets tested
• The memory pruning module is independently validated: on LongMemEval-S, generation F1 improves from 19.55 to 20.31
• Limitation: a footnote in the introduction explicitly scopes the discussion to "dialogue memory," distinct from the "experience memory" agents accumulate while executing tasks — whether the method transfers to tool-use trajectory memory is untested; the paper also provides no public code link, so external reproduction requires rebuilding from the stated hyperparameters

Reviewer's One-Line Take

The 4-dataset × 9-baseline comparison is solidly scoped, and the added theoretical upper-bound proof adds real persuasive weight; but there's no public code, and the authors themselves scope the claim to dialogue memory — readers shouldn't extrapolate these results directly onto an agent's task-execution memory.

Takeaways for You

• If you're maintaining a live conversational agent's memory system: you don't need a full redesign. Try bucketing existing memories into episodic / personal-semantic / general-semantic first, check whether different query types really do show a performance gap under your current single-strategy retrieval, then decide whether routing is worth adopting.
• If you're A/B-testing a memory system: the ablation design here — swap type-aware routing for any single strategy — gives you a ready-made control experiment to check whether your own "one-size-fits-all" retrieval strategy actually has room to improve.

Paper Two｜From Retrieval to Reconstruction: A Memory Graph That Tells Facts from Opinions

From Retrieval to Reconstruction: Constructing Evolvable Cognitive Memory for Long-Term Dialogue Zirui Liao, Zhengxian Wu, Zhuohong Chen et al. (Tsinghua University Shenzhen International Graduate School) · arxiv: 2610.11314

Links: arxiv · alphaxiv

TL;DR

A cognitive graph schema that separates "facts" from "attributed opinions," paired with a ReAct agent that decides its own next retrieval step, substantially outperforms seven baselines on multi-hop and temporal reasoning; accepted to EMNLP 2026 main conference, with multi-hop F1 dropping 25.61 points once the agentic retrieval loop is removed.

Editorial Judgment

Field Context

Standard RAG converts all text into vectors for similarity matching, and this "store-and-retrieve" paradigm carries two risks for long-term dialogue agents: semantically close content can collapse together in vector space, and the system can't adapt its search strategy based on intermediate retrieval results. The authors call the first risk "semantic collapse," and it's the core problem this paper tackles.

Mid-Level Walkthrough

• The problem: Imagine a conversation where the fact is "Jon lost his job," while someone else's opinion is "Gina thinks Jon's job loss was a mistake." Because the wording overlaps heavily, these two statements end up very close together in vector space (the authors measure an average cosine similarity of 0.8231 for matched fact-claim pairs). So when you ask "what happened to Jon," the system might hand back Gina's subjective opinion as if it were an objective fact — in the authors' controlled probe, this misattribution happened in 64 of 150 cases, 42.7% of the time.
• The method: CogMem designs a PEC²F (Person-Event-Concept-Claim-Fact) graph schema, where dedicated Claim nodes preserve "who said it" and "about whom" source information, kept separate from neutral Event/Fact nodes. Retrieval isn't a one-shot vector match — a ReAct-based cognitive search agent dynamically navigates the graph using four operators (anchoring, traversal, intersection, evidence grounding) based on query intent. During ongoing dialogue, scattered episodic records are periodically consolidated into higher-level semantic facts, reducing the number of nodes later searches need to traverse.
• Why it matters: This paper shows that "whether memory should actively probe and distinguish information sources" matters as much as "how memory is stored." Removing the agentic retrieval loop and replacing it with a fixed two-hop traversal causes multi-hop F1 to collapse by nearly half (48.42% → 22.81%) — meaning whether retrieval can "dynamically decide its next move" is itself the single most impactful design choice, more so than how memory is classified.

Key Findings

• Evaluated on 2 benchmarks (LoCoMo, LongMemEval) with 2 backbones (GPT-4o-mini, Qwen2.5-14B), against 8 baselines (Naive RAG, LightRAG, HippoRAG, RoG, Mem0, A-MEM, LightMem, GAM)
• Under the Qwen2.5-14B backbone, multi-hop F1 reaches 48.42%, 5.46 points above the strongest memory baseline GAM (42.96%)
• Overall LongMemEval accuracy is the highest among all compared methods under both backbones (68.40% / 66.50%)
• Ablation: removing the agentic retrieval loop drops multi-hop F1 from 48.42% to 22.81%, a 25.61-point fall and the largest of the three ablations; removing Claim nodes (flattening them into generic edges) drops Knowledge Update accuracy from 73.08% to 56.67%, a 16.41-point fall
• An additional blind study with two graduate-student annotators on 50 sampled questions: CogMem wins 68% of comparisons, ties 20%, loses 12%
• Limitation: the authors explicitly note the RoG comparison is a zero-shot transfer evaluation — RoG was not fine-tuned on these datasets, so this result shouldn't be read as "learned graph traversal is always worse than agentic retrieval"; all authors share a single institution, so independent external replication remains to be seen

Reviewer's One-Line Take

Three ablations cleanly isolate the "agentic retrieval loop" as the single largest contributor to the gains, and the added blind human evaluation plus the controlled semantic-collapse probe make the methodology among the more rigorous in today's candidate pool — but the authors' own admission that the RoG comparison isn't an even-conditions contrast leaves some ambiguity about how much of the F1 gap is algorithmic advantage versus training-condition disparity.

Takeaways for You

• If your conversational agent handles multi-party dialogue where "who said what" matters: check whether your current memory system mixes facts and reported opinions in the same vector space. The semantic-collapse probe here (computing cosine similarity between fact-claim pairs) can be borrowed directly to quickly diagnose whether your system has this problem.
• If you're designing an agent's memory retrieval module: these ablation results suggest that letting the retrieval step "dynamically decide its next move" may matter more than how memory is classified and stored — worth prioritizing before you invest heavily in schema design.

Today's Takeaway

I used to think the most important thing about a memory system was storing more, and storing it more precisely. Today's two papers together show that the real fork in the road is elsewhere: whether to first classify what kind of memory this is before routing it (MemoType), or whether to actively distinguish fact from attributed opinion at retrieval time and let retrieval itself become a self-directing loop (CogMem). Both papers use ablations to prove that this "extra step of judgment" matters more than simply scaling up the vector store — and both happened to be accepted the same week, to NeurIPS 2026 and EMNLP 2026 respectively, a small signal that memory management is shifting from undifferentiated retrieval toward classify-then-decide.

References

• Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies — arXiv
• MemoType — alphaXiv
• From Retrieval to Reconstruction: Constructing Evolvable Cognitive Memory for Long-Term Dialogue — arXiv
• CogMem — alphaXiv
• CogMem — code (GitHub)