AuthorsXixuan Hao, Zeyu Zhang, Zehao Lin et al.
TL;DR
MemOps reformulates long-term conversational memory as explicit lifecycle operations, revealing that Claude-Sonnet-4.5 reaches 0.916 accuracy while Temp-LoRA collapses to 0.162.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing long-term memory benchmarks mostly score downstream QA, so MemOps argues they “credit correct answers despite their reliance on inconsistent or unsafe memory states.”
In dynamic multi-session agents, this black-box view hides whether systems miss memory-relevant facts, mishandle updates, or leak forgotten values, breaking safety and personalization.
HOW IT WORKS
MemOps defines five explicit operations — Remember, Forget, Update, Reflect, TrajectoryOps — and six probes including OperationTrace and StateTransition over structured traces.
You can think of MemOps like a debugging dashboard for memory: instead of only checking outputs, it inspects every read write update delete step in a conversation.
This operation-centric design lets MemOps expose leakage, stale-value reuse, and trajectory failures that a plain context window or vanilla QA benchmark cannot reveal.
DIAGRAM
This diagram shows how MemOps turns an evidence conversation into operation traces and six probe types for a single topic.
DIAGRAM
This diagram shows the four-stage MemOps pipeline from background profiles to adjacent and long-context evaluation.
PROCESS
Background Construction
MemOps samples topic-specific user backgrounds like Personal Profile or Plans and Events to seed lifecycle operations and conversation scenarios.
Evidence Conversation and Gold Trace Generation
MemOps builds three-segment dialogues and simultaneously annotates triggers, targets, old new values, and evidence spans into a gold operation trace.
Operation-Level Probe Generation
MemOps creates six probe types — OperationTrace, TargetBinding, StateTransition, CandidateDisambiguation, OperationApplication, StateTrajectory — with gold provenance.
Long-Context Dialogue Generation
MemOps disperses evidence segments into UltraChat histories, injects context-level distractors, then evaluates adjacent and long-context responses with an LLM judge.
KEY CONTRIBUTIONS
Lifecycle operation formulation of conversational memory
MemOps formalizes Remember, Forget, Update, Reflect, and TrajectoryOps with structured triggers, targets, scopes, and state transitions, averaging 4.0 operations per trajectory.
Controllable benchmark generation pipeline
MemOps constructs 403 evidence conversations and 2,006 QA pairs via a four-stage pipeline with schema gates and LLM verification to ensure trace integrity and leakage control.
Operation-level evaluation and findings
MemOps shows session-level RAG reaches 0.845 accuracy while Temp-LoRA drops to 0.162, and reveals persistent weaknesses in state trajectory reconstruction across strong models.
RESULTS
Accuracy (Adjacent Claude-Sonnet-4.5)
0.916
+0.066 over Temp-LoRA
Accuracy (Adjacent Qwen3.6-27B)
0.914
+0.064 over Temp-LoRA
Accuracy (Session-level RAG)
0.845
+0.227 over Turn-level RAG
Accuracy (Temp-LoRA)
0.162
Parametric memory baseline on MemOps
MemOps evaluates 4,012 instances across adjacent and long-context settings, stressing lifecycle operations rather than plain QA. These numbers show how MemOps separates robust systems like Claude-Sonnet-4.5 from fragile parametric memory such as Temp-LoRA on long-horizon conversational memory.
BENCHMARK
MemOps evaluates 4,012 instances across adjacent and long-context settings, stressing lifecycle operations rather than plain QA. These numbers show how MemOps separates robust systems like Claude-Sonnet-4.5 from fragile parametric memory such as Temp-LoRA on long-horizon conversational memory.
BENCHMARK
Answer accuracy on MemOps across representative long-context and memory systems.
BENCHMARK
Comparing long-context, retrieval-based, managed, and parametric memory on MemOps accuracy.
KEY INSIGHT
MemOps shows Temp-LoRA achieves only 0.162 accuracy despite encoding history into parameters, while session-level RAG reaches 0.845 on the same benchmark.
This is surprising because many assume parametric memory should help long-term retention, yet MemOps reveals it fails badly at controlled lifecycle operations and state transitions.
WHY IT MATTERS
MemOps unlocks auditable, operation-level diagnosis of long-term memory, letting builders see exactly where Remember, Forget, Update, Reflect, or TrajectoryOps fail.
With MemOps, practitioners can design memory controllers and agents that track safe state trajectories, minimize leakage, and reason over multi-session histories in ways previous QA-only benchmarks could not support.
Guilin Zhang, Wei Jiang et al.
· 2026
A-MAC scores candidate memories using Utility, Confidence, Novelty, Recency, and Type Prior combined by a learned linear admission policy with Algorithm 1 A-MAC Memory Admission. On the LoCoMo benchmark, A-MAC achieves F1 0.583 and 2644 ms latency, improving F1 by 0.042 and reducing latency by 1187 ms compared to A-mem.
Robbyant Team, Zelin Gao et al.
arXiv 2026 · 2026
LingBot-World combines a Data Engine, Fundamental World Model, Action-Conditioned World Model, and Post-Training causal adaptation to turn a 28B-parameter video generator into a real-time interactive world simulator. On the VBench benchmark, LingBot-World achieves a dynamic degree of 0.8857 versus 0.7612 for Yume-1.5, while also improving imaging quality to 0.6683.
Manoj Madushanka Perera, Adnan Mahmood et al.
· 2026
AgenticAI-DialogGen chains ChatPreprocessor, KnowledgeExtractor, TopicAnalyzer, KnowledgeGraphBuilder, PersonaGenerator, DuelingChat Agent, ConversationValidator, ConversationRefiner, QAGeneration, and PostProcessing to turn raw multi-session chats into topic-guided, persona-grounded conversations with explicit short- and long-term memories. On the TGC / KG memory QA benchmark, Mistral-7B fine-tuned within AgenticAI-DialogGen achieves 87.36 F1, compared to GPT-4’s 83.77 F1 in a zero-shot setting on the same task.
Answers use this explainer on Memory Papers.
Checking…