AuthorsSeungbin Yang, Chaewoon Ki, Dohyun Lee et al.
TL;DR
PACMEM in PERSONATRAIL turns raw browsing trajectories into factual and preference memories, reaching 71.79% TSR vs 46.45% for ReasoningBank on single-hop preference inference with Qwen3.6-27B (+7.16pp).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
PERSONATRAIL shows that No Retrieval agents have near-zero Task Success Rate on both preference inference and episodic grounding, even when intent scores exceed 79%.
Without structured history, web agents like GPT-5.4-mini and Qwen3.6-27B satisfy generic intent but fail to apply user preferences or recover specific episodes, breaking realistic personalization.
HOW IT WORKS
PACMEM builds Factual Memory, Preference Memory, Trajectory Segmentation, and Memory Retrieval over PERSONATRAIL’s browser-level histories to preserve episodic details and recurring habits.
You can think of Factual Memory as a diary of past browsing episodes and Preference Memory as a profile summarizing long-term tastes, like RAM versus a user manual.
This dual-memory design lets PACMEM answer queries like “movie I usually enjoy” or “page I saw last Tuesday” that a plain context window cannot reliably ground.
DIAGRAM
This diagram shows how PACMEM uses factual and preference memories during inference to answer preference inference and episodic grounding queries.
DIAGRAM
This diagram shows the PERSONATRAIL pipeline from website curation to personalized query generation and verification.
PROCESS
Trajectory Segmentation
PACMEM segments raw browsing logs into goal-oriented chunks, preparing inputs for Factual Memory Construction and Preference Memory Construction while preserving timestamps and URLs.
Factual Memory Construction
PACMEM groups related segments into episodes using embeddings and an LLM, then stores structured entries in Factual Memory with titles and interaction sequences.
Preference Memory Construction
PACMEM builds a similarity graph over factual entries, clusters them, and uses an LLM to derive Preference Memory summaries like preferred genres or sites.
Memory Retrieval
At query time, PACMEM retrieves and reranks top-K factual and preference entries, then feeds them to the web agent to guide personalized navigation.
KEY CONTRIBUTIONS
PERSONATRAIL benchmark for personalized web agents
PERSONATRAIL offers 2,524 queries over 317 websites and 100 users, with browser-level trajectories and both preference inference and episodic grounding tasks.
PACMEM dual memory framework
PACMEM introduces Factual Memory and Preference Memory built via Trajectory Segmentation and clustering, providing a structured baseline for personalized web navigation.
Comprehensive evaluation on PERSONATRAIL
PACMEM is evaluated against AWM and ReasoningBank across six backbones, achieving up to 71.79% TSR on single-hop preference inference with Qwen3.6-27B.
RESULTS
Task Success Rate
53.61%
+7.16 over ReasoningBank on single-hop preference inference with Qwen3.6-27B
Task Success Rate
39.03%
+21.19 over AWM on multi-hop episodic grounding with Qwen3.6-27B
Preference Score
80.46%
single-hop preference inference with Claude Haiku 4.5, vs 62.51 for ReasoningBank
Episodic Retrieval Recall
96.77%
single-hop episodic grounding with Claude Haiku 4.5, vs 63.07 for ReasoningBank
These results are on the PERSONATRAIL benchmark, which tests preference inference and episodic grounding. The gains show PACMEM’s dual-memory design improves both personalization accuracy and exact episodic retrieval over memory-augmented baselines.
BENCHMARK
These results are on the PERSONATRAIL benchmark, which tests preference inference and episodic grounding. The gains show PACMEM’s dual-memory design improves both personalization accuracy and exact episodic retrieval over memory-augmented baselines.
BENCHMARK
Task Success Rate on single-hop preference inference for Qwen3.6-27B across memory systems.
BENCHMARK
Aggregated Task Success Rate over single-hop and multi-hop tasks for Gemini-3-Flash under different PACMEM memory configurations.
KEY INSIGHT
Even strong backbones like GPT-5.4-mini reach 79.81% Intent Score without retrieval, yet achieve 0.00% Task Success Rate on preference inference.
This is surprising because high intent satisfaction suggests good navigation, but PERSONATRAIL reveals that without PACMEM-style memory, agents almost never satisfy personalization constraints.
WHY IT MATTERS
PACMEM shows that structuring browsing history into factual and preference memories enables robust preference inference and episodic grounding across diverse web tasks.
Builders can now design web agents that remember user habits and specific past pages, making realistic long-term personalization feasible in managed open web environments.
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…