AuthorsZhiyu Shen, Ziming Wu, Fuming Lai et al.
TL;DR
MemBuilder uses Attributed Dense Rewards Policy Optimization with multi-dimensional memory to reach 84.23% on LoCoMo, +6.75 points over MIRIX.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Long-term dialogue agents often receive a single reward at trajectory end, leaving MemBuilder unable to tell which memory operations helped or hurt.
This sparse feedback breaks credit assignment across dozens of sessions, so multi-session QA and temporal reasoning remain unreliable despite large context windows.
HOW IT WORKS
MemBuilder combines Multi-Dimensional Memory Architecture, Supervised Fine-Tuning, and Attributed Dense Rewards Policy Optimization so Qwen3-4B can manage Core, Episodic, Semantic, and Procedural memory.
Think of MemBuilder as a hybrid between RAM and a card catalog, where dense questions probe each session’s memory and gradients are scaled by how often components are retrieved.
This design lets MemBuilder build structured external memory that supports cross-session, temporal, and adversarial QA far beyond what a plain context window or naive RAG can handle.
DIAGRAM
This diagram shows how MemBuilder uses synthetic session-level QA and Attributed Dense Rewards Policy Optimization to train memory construction.
DIAGRAM
This diagram shows how MemBuilder is trained on LongMemEval and evaluated on LoCoMo and PerLTQA with fixed answer models and baselines.
PROCESS
Multi-Dimensional Memory Architecture
MemBuilder decomposes the dialogue into Core Memory, Episodic Memory, Semantic Memory, and Procedural Memory, structuring events, facts, and routines for later retrieval.
Supervised Fine-Tuning
MemBuilder uses expert trajectories from Claude 4.5 Sonnet to learn valid JSON actions and operation types for each memory component before reinforcement learning.
Attributed Dense Rewards Policy Optimization
MemBuilder applies ADRPO with synthetic session-level QA, format validity gating, and length penalties to update memory policies using dense rewards.
Contribution-Aware Gradient Weighting
MemBuilder scales gradients for Episodic, Semantic, and Procedural Memory based on retrieval counts, amplifying the dominant contributor while keeping Core Memory weight fixed.
KEY CONTRIBUTIONS
Multi-Dimensional Memory Architecture
MemBuilder defines Core Memory, Episodic Memory, Semantic Memory, and Procedural Memory, enabling a 4B model to reach 84.23% on LoCoMo with structured external memory.
Attributed Dense Rewards Policy Optimization
MemBuilder introduces ADRPO with synthetic session-level QA, format validity gating, and length penalties to replace sparse trajectory rewards in long-term dialogue training.
Contribution-Aware Gradient Weighting
MemBuilder weights gradients by retrieval-based contribution, using α = 4 to raise LoCoMo accuracy from 82.82% to 84.04% in ablations.
RESULTS
LoCoMo accuracy
84.23%
+6.75 over MIRIX
LongMemEval accuracy
85.75%
+12.75 over RAG Session
PerLTQA accuracy
93.14%
+13.93 over RAG Session
Parameter count
4B
Qwen3 4B Instruct base size for MemBuilder
MemBuilder is trained only on LongMemEval yet evaluated on LoCoMo and PerLTQA, which test multi-session QA, temporal reasoning, and evolving user profiles. The 84.23% LoCoMo result shows MemBuilder can beat prompting frameworks using Claude 4.5 Sonnet while running on a 4B open-source backbone.
BENCHMARK
MemBuilder is trained only on LongMemEval yet evaluated on LoCoMo and PerLTQA, which test multi-session QA, temporal reasoning, and evolving user profiles. The 84.23% LoCoMo result shows MemBuilder can beat prompting frameworks using Claude 4.5 Sonnet while running on a 4B open-source backbone.
BENCHMARK
Accuracy on LoCoMo multi-session dialogue QA.
KEY INSIGHT
MemBuilder trained with both SFT and RL reaches 84.23% on LoCoMo, while RL alone drops to 79.31%, below the 81.74% SFT-only baseline.
This is surprising because many assume more RL always helps, but MemBuilder shows that dense rewards without a strong supervised cold-start can actually hurt performance.
WHY IT MATTERS
MemBuilder proves that a 4B open-source model can handle long-term memory construction with dense rewards and multi-dimensional attribution, rivaling closed-source prompting systems.
Builders can now deploy cost-effective agents that maintain structured, evolving memories across thousands of tokens, enabling robust multi-session QA and temporal reasoning without relying on expensive APIs.
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…