AuthorsLin Du, Jie Zhou, Yuxuan Cai et al.
TL;DR
AutoMem uses text-gradient guided Experience-Guided Architecture Search and Failure-Guided Module Diagnosis to find task-adaptive memory architectures that reach 71.5% on GAIA vs 67.8% for MemoryBank (+3.7 points).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
AutoMem’s pilot random search shows the best sampled memory architecture reaches 69.7% on GAIA, while the strongest fixed baseline MemoryBank reaches 67.8%.
The same architecture that scores 69.7% on GAIA can underperform the no-memory agent on other benchmarks, exposing task-specific, module-coupled memory failures.
HOW IT WORKS
AutoMem’s core mechanism factorizes memory into Encode, Store, Retrieve, and Manage, and optimizes them via Experience-Guided Architecture Search and Failure-Guided Module Diagnosis.
You can think of AutoMem like a modular computer: Encode is the CPU writing instructions, Store is disk format, Retrieve is the index, and Manage is the garbage collector.
This text-gradient design lets AutoMem rewire memory architectures themselves, enabling improvements beyond what any fixed context window or single retrieval component can provide.
DIAGRAM
This diagram shows how AutoMem runs one text-gradient search round from proposal to evaluation and diagnosis.
DIAGRAM
This diagram shows how AutoMem splits data and compares discovered architectures against baselines on GAIA, WebWalkerQA, and xBench-DeepSearch.
PROCESS
Experience-Guided Architecture Search
AutoMem uses Experience-Guided Architecture Search to propose K candidate tuples over Encode, Store, Retrieve, and Manage, conditioned on the Pareto front and experience ledger.
Agent Rollouts on Search Batch
AutoMem instantiates each candidate architecture as a memory agent Aa and runs full rollouts on the search batch Bs to measure accuracy, memory-induced gain, and cost.
Pareto Front Update and Validation
AutoMem updates the Pareto front Pt using non-dominated architectures and periodically validates the current leader on the held-out validation batch Bv.
Failure-Guided Module Diagnosis
AutoMem applies Failure-Guided Module Diagnosis to failed rollouts, attributing failures to Encode, Store, Retrieve, or Manage and synthesizing a textual gradient δ for the next round.
KEY CONTRIBUTIONS
Modular View of Agent Memory Architectures
AutoMem systematizes long-term memory as four modules Encode, Store, Retrieve, and Manage, and defines a feasible architecture space A with 5 encoders, 5 stores, 6 retrievers, and 4 managers.
Text-Gradient Recursive Self-Improvement
AutoMem introduces a text-gradient loop combining Experience-Guided Architecture Search and Failure-Guided Module Diagnosis to iteratively refine memory architectures under limited evaluation budgets.
Task-Adaptive Memory Architectures on GAIA WebWalkerQA xBench
AutoMem discovers task-adaptive architectures that reach 71.5% on GAIA, 72.5% on WebWalkerQA, and 46.0% on xBench-DeepSearch, improving accuracy by 2.8 points on average across six benchmark-backbone settings.
RESULTS
GAIA Perf.
71.5%
+3.7 over MemoryBank
GAIA Cost
128.4k tokens
−72.0k vs MemoryBank
WebWalkerQA Perf.
72.5%
+3.6 over Agent-KB
xBench Perf.
46.0%
+1.0 over MemoryBank
On GAIA, WebWalkerQA, and xBench-DeepSearch with Qwen3.5-122B-A10B, AutoMem is evaluated against strong memory baselines like MemoryBank and Agent-KB. The main result shows AutoMem finds architectures that increase GAIA accuracy from 67.8% to 71.5% while cutting GAIA token cost from 200.4k to 128.4k.
BENCHMARK
On GAIA, WebWalkerQA, and xBench-DeepSearch with Qwen3.5-122B-A10B, AutoMem is evaluated against strong memory baselines like MemoryBank and Agent-KB. The main result shows AutoMem finds architectures that increase GAIA accuracy from 67.8% to 71.5% while cutting GAIA token cost from 200.4k to 128.4k.
BENCHMARK
Accuracy (%) on GAIA comparing AutoMem to memory baselines from Table 2.
KEY INSIGHT
AutoMem reaches 71.5% on GAIA after only five guided rounds, beating the best of ten random architectures at 69.7% while using 0.55× the token budget.
This is surprising because random search already found architectures surpassing MemoryBank, yet AutoMem’s text-gradient guidance still extracts another +1.8 points over the random best with fewer evaluations.
WHY IT MATTERS
AutoMem unlocks a practical way to treat long-term memory as an optimizable architecture, not a fixed add-on, using text gradients over Encode, Store, Retrieve, and Manage.
Builders can now automatically discover task-specific memory stacks for GAIA-like web agents or deep-search tasks, instead of hand-tuning a single retriever or store that may underperform across benchmarks.
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…