Research map / Multi-agent systems and games / LLM agents

Rethinking How to Remember: Beyond Atomic Facts in Lifelong LLM Agent Memory

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 0 citations (OpenAlex)

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, an LLM agent memory system keeping three granularities (raw dialogue segments linked by source IDs, atomic facts for retrieval, and person profiles) with TextGrad-optimized extraction and profile prompts, evaluated on LoCoMo and PerLTQA.

Problem

Fact-centric agent memory systems lose fine-grained details during extraction, support only shallow reasoning over scattered facts, and use static hand-written extraction prompts that yield inconsistent granularity across diverse dialogues.

Approach

Maintain three coexisting representations per memory: verbatim dialogue windows reached through source-ID pointers from each fact, atomic facts for similarity retrieval, and incrementally synthesized person profiles; extraction and profile prompts are optimized with TextGrad from answer-quality feedback, with no parameter updates.

Key results

• On LoCoMo with GPT-4.1-mini, TriMem reaches 54.26 average F1 versus 50.30 for the strongest baseline SimpleMem, using about 1.2k retrieved tokens.
• With Llama-3.1-8B-Instruct on LoCoMo, TriMem averages 38.70 F1 versus 33.18 for A-Mem.
• On PerLTQA with Qwen3-8B, TriMem scores 92.46 on Profile and 55.79 on Dialogues, the best across the four subtasks reported.

Surprising findings

• Static fact-only retrieval loses about 14.5% more reference-answer tokens than the original dialogue, per the paper's storage completeness analysis.

Limitations

• Profile construction is limited to person-level entities; higher-order relational structures are not explored.
• TextGrad optimization uses only response-level feedback, without intermediate retrieval or reasoning-step supervision.
• Prompt optimization is run on LoCoMo with Qwen3-8B and transferred to other models without per-model tuning.

Experimental setup

Datasets: LoCoMo PerLTQA

Benchmarks: LoCoMo PerLTQA

Metrics: BLEU F1 LLM-judged correctness token cost Hit@k

Baselines: Full-Context Naive RAG Mem0 MemoryOS A-Mem LightMem SimpleMem xMemory

Builds on: TextGrad Mem0 SimpleMem LightMem xMemory

Closest papers by meaning

• Memory-T1: Reinforcement Learning for Temporal Reasoning in Multi-session Agents (ICLR 2026 · similarity 0.90)
• PlugMem: A Task-Agnostic Plugin Memory Module for LLM Agents (ICML 2026 · similarity 0.89)
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (ICML 2026 · similarity 0.89)
• Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.88)
• StateTree: Enhancing Long-term Dialogue Reasoning via Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
• SimpleMem: Efficient Lifelong Memory for LLM Agents (ICML 2026 · similarity 0.87)