Tag
The paper proposes T-Mem, a temporal-structure-based memory framework for LLM agents in long-horizon multi-party conversations, using Temporal Evidence Discovery and Temporal Memory Structuring to recover locally concentrated evidence and reconstruct event progressions as an evidence forest, outperforming existing memory frameworks on two benchmarks.
TWIST is a proposed benchmark suite for evaluating intervention quality in conversational memory systems, featuring human-validated tracks to detect failures like unresolved tensions and stale facts, revealing trade-offs between recall and specificity that traditional metrics miss.
The paper introduces LSREP, a longitudinal state-replay protocol for evaluating conversational memory, with ICE v2 as a case study, revealing failure modes through replay and auditing in comparison to vector-RAG.
This paper introduces SCALE-QA, a benchmark for evaluating interleaved conversational memory in long contexts, and proposes TSIM, a method that improves memory reconstruction accuracy over existing baselines.
This paper introduces ReFind, an agent-controlled search interface that works over raw, unmodified chat logs using lexical indexing and iterative keyword search, showing it can rival structured memory systems like HippoRAG 2 on conversational-memory benchmarks without building any semantic structure.
Introduces MemOps, a benchmark that reformulates conversational memory as lifecycle operations with structured traces, enabling operation-level diagnosis of memory failures in LLM-based agents, revealing that current systems are far from uniformly reliable.
ConvMemory v3 introduces a validity context layer that detects outdated or superseded conversational memories using a target-conditioned dual-evidence gate, achieving high accuracy on synthetic benchmarks and zero-shot transfer to role binding tasks.
ConvMemory v2 is a recall-preserving reranker that reorders the top-10 candidates from ConvMemory v1 using a fine-tuned cross-encoder, improving MRR on the LoCoMo benchmark while preserving recall.
MemoryDocDataSet is a new synthetic benchmark of 50 micro-worlds and 1,000 QA pairs designed to evaluate AI systems on the joint task of conversational memory and long-document reasoning simultaneously. The best baseline (RAG-Both) achieves only 0.358 overall F1, highlighting a significant gap in current systems' ability to unify conversational memory with long-document navigation.
A novel memory retrieval system inspired by episodic memory theory achieves state-of-the-art 96.4% top-50 accuracy on the LongMemEval benchmark using Gemini Flash, outperforming larger Pro-based baselines by isolating retrieval quality from model capability.
EviMem combines IRIS for evidence-gap detection and LaceMem for layered memory to improve long-term conversational memory retrieval, achieving higher accuracy on temporal and multi-hop questions with lower latency.