AuthorsZhaopei Huang, Qifeng Dai, Guozheng Wu et al.
TL;DR
H2Memory in Mem-PAL builds hierarchical heterogeneous memories over logs and dialogues, reaching a GPT-4 G-Score of 32.54 vs 29.33 for Session-level RAG on PAL-Bench.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing service-oriented assistants often overlook long-term interactions and fail to capture users’ subjective characteristics and preferences over time.
This causes user-agent systems to misinterpret implicit needs and personalized requirements, forcing users to repeat context and reducing preference-aligned responses.
HOW IT WORKS
Mem-PAL’s core mechanism is H2Memory, combining Log Graph, Background, Topic Outline, and Principle memories with retrieval-augmented generation over long-term histories.
Think of Log Graph and Topic Outline as fast-access notes, while Background and Principle act like a slowly updated card catalog of the user’s life and preferences.
This hierarchical heterogeneous design lets Mem-PAL infer implicit requirements and stable preference principles that a plain context window or flat memory cannot represent.
DIAGRAM
This diagram shows how Mem-PAL retrieves heterogeneous memory entries from H2Memory and uses them to generate a personalized response for a current query.
DIAGRAM
This diagram shows how Mem-PAL’s PAL-Set is synthesized and then used in PAL-Bench’s three evaluation tasks.
PROCESS
Log Graph Construction
Mem-PAL builds Log Graph by linking fragmented logs with Caused by and Follows relations, forming connected subgraphs and situation descriptions for each session.
Background Memory Updating
Mem-PAL recursively updates Background memory by summarizing situation entries across sessions into aspect-based paragraphs, keeping long-term user context current.
Topic Outline Extraction and Requirement Rewriting
Mem-PAL segments dialogues into topics, builds Topic Outline schemas, and rewrites requirements using retrieved situations from Log Graph to enrich underspecified queries.
Principle Clustering and Personalized RAG
Mem-PAL clusters requirements to form Principle memory, then retrieves entries from all components of H2Memory to drive personalized retrieval-augmented generation for responses.
KEY CONTRIBUTIONS
PAL-Bench and PAL-Set Construction
Mem-PAL introduces PAL-Bench and PAL-Set with 100 users, averaging 29 sessions, 996 logs, and 401 dialogue turns per user, synthesized via a multi-stage pipeline.
H2Memory Hierarchical Heterogeneous Memory
Mem-PAL’s H2Memory combines Log Graph, Background, Topic Outline, and Principle memories with RAG to model both concrete events and abstract preferences.
Comprehensive Evaluation on PAL-Bench and LongMemEval
Mem-PAL demonstrates gains on PAL-Bench tasks and reaches 50.00% accuracy on LongMemEval’s single-session-preference subset using Topic Outline and Principle memories.
RESULTS
G-Score Requirement Restatement
32.54
+3.21 over Session-level RAG
BLEU-1 Requirement Restatement
26.67
+2.86 over MemoryBank
S-Score Solution Proposal
38.32
+4.54 over MemoryBank
Win Count Requirement Multi-turn
478
vs 29 ties and 319 losses against Vanilla w/o log
On PAL-Bench, Mem-PAL’s H2Memory is evaluated on Requirement Restatement, Solution Proposal, and Multi-turn Dialogue Interaction. These results show Mem-PAL’s hierarchical heterogeneous memory improves requirement inference and preference-aligned solutions compared to Session-level RAG and MemoryBank.
BENCHMARK
On PAL-Bench, Mem-PAL’s H2Memory is evaluated on Requirement Restatement, Solution Proposal, and Multi-turn Dialogue Interaction. These results show Mem-PAL’s hierarchical heterogeneous memory improves requirement inference and preference-aligned solutions compared to Session-level RAG and MemoryBank.
BENCHMARK
GPT-4-based G-Score on PAL-Bench Requirement Restatement task.
BENCHMARK
Selection Score on PAL-Bench Solution Proposal task.
KEY INSIGHT
Removing Topic Outline memory drops Solution Proposal S-Score from 38.32 to 28.09, a 10.23 decrease, far larger than removing other components.
This is surprising because one might expect abstract Principle memory to dominate preference modeling, but Mem-PAL shows structured topic-level histories are more critical.
WHY IT MATTERS
Mem-PAL unlocks assistants that can jointly reason over behavioral logs and dialogues to infer implicit requirements and nuanced preference principles.
Builders can now design long-term personalized dialogue agents that maintain evolving personas across months of interaction without retraining, using H2Memory as an external memory backbone.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…