AuthorsWenya Xie, Shengming Zhou, Zelin Li et al.
TL;DR
DynamicMem uses multi-timescale user profile evolution with intent-conditioned event chains to expose that State Completion can drop by up to 26.5 points over 15 months while Personalized Service stays flat.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
DynamicMem shows that existing memory systems cannot reliably track a user’s evolving attributes, habits, and preferences over 15 months and ~2.2M tokens per user.
When State Completion is evaluated at five checkpoints, accuracy steadily degrades while Personalized Service stays flat, masking failures in retaining stable facts and overwriting changed ones.
HOW IT WORKS
DynamicMem’s core mechanism is a three-stage pipeline: Multi-timescale user profile construction, Intent-conditioned event chain generation, and State-consistent multi-app log generation that together produce causally grounded trajectories.
You can think of DynamicMem like a long-running operating system log: quarterly world background updates change the “configuration,” and event chains are the processes whose API calls mutate persistent app state.
This design lets DynamicMem probe what long-horizon memory architectures can do beyond a plain context window, including tracking typed deltas, reconciling conflicting evidence, and inferring implicit preferences from scattered behavior.
DIAGRAM
This diagram shows how DynamicMem evaluates memory systems at quarterly checkpoints using State Completion and Personalized Service tasks.
DIAGRAM
This diagram shows how DynamicMem evolves attributes, habits, and preferences over five quarterly windows driven by world background.
PROCESS
Multi-timescale user profile construction
DynamicMem first builds a base profile and quarterly dynamic state over six life domains, separating attributes, habits, and preferences with typed deltas like ADD, ACQUIRE, and SHIFT.
Intent-conditioned event chain generation
DynamicMem then turns each state change or stable item into event chains across 16 apps, where each step has a fine-grained intent and app–API pair reflecting the underlying profile.
State-consistent multi-app log generation
DynamicMem grounds event chains into concrete API calls and responses, maintaining stateful app environments so carts, playlists, and threads stay consistent across 17,715 app-log events.
Checkpoint-based evaluation
DynamicMem defines five quarterly checkpoints and, at each, runs State Completion and Personalized Service tasks over 3,634 problems to see how memory systems scale with growing history.
KEY CONTRIBUTIONS
Long-horizon multi-app trajectories
DynamicMem generates 15-month trajectories with ~2.2M tokens per user, 1,790 event chains, and 17,715 grounded events across 16 applications and 66 APIs.
Multi-timescale user modeling
DynamicMem decomposes user state into attributes, habits, and preferences, tracking 883 attribute, 466 habit, and 488 preference instances with 383 typed state transitions.
Checkpoint-based memory diagnosis
DynamicMem introduces five quarterly checkpoints and two task families, yielding 1,824 State Completion and 1,810 Personalized Service problems that reveal distinct retention and update failures.
RESULTS
State Completion drop A-Mem
26.5 points
-26.5 vs C1 for A-Mem Preferences
State Completion drop RAG
4.4 points
-4.4 vs C1 overall State Completion
Personalized Service change RAG
+2.8 points
+2.8 vs C1 overall Personalized Service
Failure due to memory
93%
93% of failures traced to memory retrieval
DynamicMem evaluates Vanilla RAG, HippoRAG2, A-Mem, MemoryOS, and SimpleMem on 3,634 checkpoint-based problems. The main result shows State Completion can decline by up to 26.5 points over five checkpoints while Personalized Service remains stable, and over 93% of failures are caused by memory systems rather than answer-generation LLMs.
BENCHMARK
DynamicMem evaluates Vanilla RAG, HippoRAG2, A-Mem, MemoryOS, and SimpleMem on 3,634 checkpoint-based problems. The main result shows State Completion can decline by up to 26.5 points over five checkpoints while Personalized Service remains stable, and over 93% of failures are caused by memory systems rather than answer-generation LLMs.
BENCHMARK
Score drop on Preference State Completion between first and fifth checkpoints.
KEY INSIGHT
DynamicMem reveals that as trajectories grow to 15 months, State Completion steadily degrades while Personalized Service scores remain flat or even improve slightly.
This is counterintuitive because both tasks read the same memory, yet longer histories help scenario-based assistance while hurting direct state reconstruction, breaking the assumption that more evidence uniformly helps.
WHY IT MATTERS
DynamicMem unlocks a controlled way to stress-test long-horizon memory architectures on realistic, evolving user behavior across multiple applications and life domains.
Builders can now design and compare memory systems that separately optimize retention and update, diagnose identity versus detail failures, and tune mechanisms for implicit preferences and proactive habit selection.
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…