AuthorsYubo Wang, Qiuyu Zhao, Zenghui Sun et al.
TL;DR
CMI-Mem uses an action-conditioned Conditional Mutual Information reward to train long-term memory management, reaching 72.1% on LoCoMo vs 68.1% for MemBuilder-RL-4B (+4.0 points).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing QA-based memory managers value memory only through sampled questions and a fixed reader, leaving many useful facts unsupervised.
This query-conditioned supervision causes distribution mismatch and coarse credit assignment, so long-term personalization and cross-session reasoning degrade when future queries differ from synthetic QA pairs.
HOW IT WORKS
CMI-Mem combines a Structured Multi-dimensional Memory Architecture, Action-conditioned CMI Reward, and Reinforcement Training with GRPO to manage long-term dialogue memory across four cognitive slots.
You can think of CMI-Mem as a smart cache plus librarian: it treats memory like RAM and disk, deciding which fragments are novel, relevant, or redundant.
By scoring each operation with Conditional Mutual Information, CMI-Mem preserves reusable information beyond a plain context window while still being anchored by QA correctness.
DIAGRAM
This diagram shows how CMI-Mem processes each dialogue session, updates memory slots, and computes per session CMI rewards.
DIAGRAM
This diagram shows how CMI-Mem rolls out sessions, collects QA and CMI rewards, and updates the policy with GRPO.
PROCESS
Structured Multi dimensional Memory Architecture
CMI-Mem receives session Ct and routes it into the Structured Multi-dimensional Memory Architecture, exposing Core, Episodic, Semantic, and Procedural slots to the shared policy.
Action conditioned CMI Reward
For each candidate fragment mnew t and local memory Mlocal t minus1, CMI-Mem computes the Action-conditioned CMI Reward I(Ct; mnew t | Mlocal t minus1).
Hybrid Reward Composition
CMI-Mem mixes the shaped CMI reward with QA correctness using rt = α rCMI t plus (1 minus α) rQA t to value each session.
Reinforcement Training with GRPO
Using Group Relative Policy Optimization, CMI-Mem updates the shared policy πθ over sessions, guided by the hybrid reward and curriculum learning.
KEY CONTRIBUTIONS
Structured Multi dimensional Memory Architecture
CMI-Mem partitions memory Mt into core, episodic, semantic, and procedural slots with type specific agents, enabling 72.1% accuracy on LoCoMo versus 50.5% for RAG top 5.
Action conditioned CMI Reward
CMI-Mem introduces an intrinsic Conditional Mutual Information reward I(Ct; mnew t | Mlocal t minus1) that scores each memory operation without conditioning on sampled QA queries.
Reinforcement Training with GRPO
CMI-Mem trains with Group Relative Policy Optimization on LongMemEval, combining CMI and QA rewards to reach 67.0% on LongMemEval versus 56.5% for MemBuilder-RL-4B.
RESULTS
LoCoMo accuracy
72.1%
+4.0 over MemBuilder-RL-4B
LME S accuracy
67.0%
+10.5 over MemBuilder-RL-4B
MABench overall score
46.2
+11.1 over MemBuilder-RL-4B
MABench overall score 8B
56.7
+5.0 over MemBuilder-RL-8B
These metrics come from LoCoMo, LongMemEval s_cleaned, and MemoryAgentBench, which test long context recall, cross session reasoning, and selective forgetting. The gains show that CMI-Mem transfers better across memory use scenarios than QA only RL baselines like MemBuilder-RL.
BENCHMARK
These metrics come from LoCoMo, LongMemEval s_cleaned, and MemoryAgentBench, which test long context recall, cross session reasoning, and selective forgetting. The gains show that CMI-Mem transfers better across memory use scenarios than QA only RL baselines like MemBuilder-RL.
BENCHMARK
Accuracy or overall score comparing CMI-Mem-4B against RAG and MemBuilder-RL-4B across three benchmarks.
KEY INSIGHT
CMI-Mem with CMI plus QA reward reaches 68.50% on LongMemEval, while CMI only achieves 45.00% and QA only reaches 56.50%.
This is surprising because intrinsic CMI supervision alone underperforms, yet when combined with QA it adds +12.0 points over QA only, contradicting the intuition that QA reward already captures all useful signal.
WHY IT MATTERS
CMI-Mem enables memory managers that optimize each operation using information theoretic CMI while still being grounded by downstream QA correctness.
Builders can now train compact memory managers that generalize across recall, temporal reasoning, and selective forgetting tasks without hand crafted rules or massive synthetic QA coverage.
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…