Research map / Large language models / Retrieval and knowledge

Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies

NeurIPS 2026 · Large language models › Retrieval and knowledge · approach: Long-context and personalized retrieval

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TriMEM, a 6,000-sample dialogue memory type dataset (episodic, personal semantic, general semantic), and MemoType, a memory framework that routes memories and queries to type-specific retrieval strategies, with a theoretical upper bound on single-strategy retrieval alignment.

Problem

Retrieval-based LLM memory systems apply one unified retrieval strategy to all memories, despite memories differing in structure and content; precise memory-type classification is also hard in topic-rich, boundary-blurred dialogue.

Approach

Categorize dialogue memories by Tulving-inspired types (episodic, personal semantic, general semantic), train a LoRA-tuned router to classify memories and queries (via hypothetical memories), and apply tailored retrieval (element decomposition, hypothetical queries, hypothetical memory) per type, plus LLM-based memory pruning.

Key results

• On LongMemEval-S, MemoType reaches Recall@1 59.15, versus 56.99 for the strongest baseline LightMem.
• On PerLTQA, MemoType reaches Recall@1 71.92 versus 57.19 for MemoryOS.
• The fine-tuned router achieves 92.10% overall accuracy on TriMEM, above GPT-4o-Mini's 83.55%.

Surprising findings

• Even lightweight encoders like BERT reportedly achieve competitive routing performance, suggesting dataset quality matters more than model scale.
• Prompting-based LLM classifiers reach only 71.9-83.6% accuracy on memory type classification, showing difficulty of the task.

Limitations

• Theoretical bound relies on an assumption of non-identical centroid directions and uniform category distribution with a shared encoder.
• Human verification covered only 20% of the dataset and the dataset is built largely from GPT-5.2 labels.
• Several evaluation baselines (A-Mem, HippoRAG2, RAPTOR) were not reported on LongMemEval-M due to construction latency.

Experimental setup

Datasets: LongMemEval-S LoCoMo PersonaMem PerLTQA

Benchmarks: LongMemEval-S LoCoMo LongMemEval-M PerLTQA

Metrics: Recall@1 Recall@3 Recall@5 Recall@10 NDCG@k GPT4Judge F1 BLEU BERTScore ROUGE

Baselines: HyDE Mill Query2Doc SeCom A-Mem HippoRAG2 RAPTOR MemoryOS LightMem

Builds on: Tulving episodic/semantic memory theory HyDE Reciprocal Rank Fusion

Closest papers by meaning

• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.88)
• A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.87)
• UMEM: Unified Memory Extraction and Management Framework for Generalizable Memory (ICML 2026 · similarity 0.87)
• Peripheral Memory for LLMs: Integration of Sequential Memory Banks with Adaptive Querying (ICML 2025 · similarity 0.86)
• Frozen Memory Is Not Enough: Rethinking External Memory as Extraction (NeurIPS 2026 · similarity 0.86)
• Memory Decoder: A Pretrained, Plug-and-Play Memory for Large Language Models (NeurIPS 2025 · similarity 0.86)