AuthorsRuiyang Jiang, Chunyu Li, Zhiyu Li
TL;DR
MemRetriever uses a ReAct-style search agent with parallel search, serial search, and reflection-based denoising to boost LOCOMO llm-as-judge accuracy from 0.785 to 0.830.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Static top k retrieval issues a single query, returns a fixed number of snippets, and often misses evidence scattered across sessions.
For long term agent memory tasks like LOCOMO, this causes incomplete recall, noisy context, and lower answer accuracy on complex multi hop questions.
HOW IT WORKS
MemRetriever builds a ReAct style search agent over a long term memory store using parallel_search, serial_search, and reflection_and_denoise tools plus a GRPO trained search policy.
You can think of MemRetriever like a smart card catalog that repeatedly checks the shelves, pulls new cards, and discards duplicates before handing a clean stack to the reader.
This KEY_MECHANISM lets MemRetriever plan when to search, how to filter evidence, and when to stop, instead of relying on a passive context window.
DIAGRAM
This diagram shows how MemRetriever runs a think action observation loop to search, reflect, and stop before answering.
DIAGRAM
This diagram shows how MemRetriever uses supervised warm start and GRPO with search, reflection, stop, and format rewards.
PROCESS
ReAct Style Search Trajectory
MemRetriever observes the question, evidence pool, and history, then starts a ReAct style trajectory with internal reasoning traces zt.
parallel_search Action
MemRetriever issues multiple queries via parallel_search when the evidence space is ambiguous, expanding recall from different retrieval directions.
serial_search Action
MemRetriever switches to serial_search once a specific information gap is identified, sending a targeted query to fill that missing fact.
reflection_and_denoise and Stop
MemRetriever runs reflection_and_denoise to filter and deduplicate evidence, then uses the done flag to decide when to stop and return the final pool.
KEY CONTRIBUTIONS
Multi Step Agentic Search for Memory Retrieval
MemRetriever formulates long term memory retrieval as a multi step agentic search process with parallel_search, serial_search, and reflection_and_denoise operations over a memory store.
Hierarchical Reward Design for Retrieval
MemRetriever uses coverage gain search rewards, recall preserving reflection rewards, and answer sufficiency stop rewards, with λdrop=0.6 and λbonus=0.1 controlling reflection penalties and bonuses.
Storage Agnostic Retrieval Decision Layer
MemRetriever models storage agnostic retrieval decision logic that can sit between agents and MemOS or Milvus, improving LOCOMO llm-as-judge accuracy from 0.785 to 0.830.
RESULTS
llm-as-judge
0.830
+0.025 over MemRetriever-4B-SFT on LOCOMO
hit@5_all
0.826
+0.034 over MemRetriever-4B-SFT on LongMemEval
token
662
40 fewer tokens than MemRetriever-4B-SFT on LongMemEval
EM
0.540
+0.040 EM over MemRetriever-4B-SFT on HotpotQA
On LOCOMO and LongMemEval, MemRetriever improves retrieval metrics and llm-as-judge accuracy while reducing token usage. On HotpotQA, MuSiQue, and 2WikiMultiHopQA, MemRetriever raises EM and F1, showing that the search agent directly benefits multi hop reasoning.
BENCHMARK
On LOCOMO and LongMemEval, MemRetriever improves retrieval metrics and llm-as-judge accuracy while reducing token usage. On HotpotQA, MuSiQue, and 2WikiMultiHopQA, MemRetriever raises EM and F1, showing that the search agent directly benefits multi hop reasoning.
BENCHMARK
llm-as-judge accuracy on LOCOMO under the same three tool retrieval pipeline.
KEY INSIGHT
MemRetriever-4B-RL reduces average tokens on LOCOMO from 754 to 673 while increasing llm-as-judge accuracy from 0.785 to 0.830.
This is counterintuitive because more retrieved memories usually seem safer, yet MemRetriever shows that reflection based denoising and stopping earlier can improve accuracy.
WHY IT MATTERS
MemRetriever unlocks a reusable retrieval decision layer that learns when to search, how to search, how to reflect, and when to stop.
Builders can now plug MemRetriever between agents and memory or vector stores to get compact, task aligned evidence pools instead of static top k dumps.
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…