Long-Horizon Memory Benchmark

DynamicMem: A Long-Horizon Memory Benchmark in Real-World Settings

Can memory systems track who a user is, what they do, and what they prefer — and keep it current — across months of multi-app activity?

Wenya Xie1, Shengming Zhou1, Zelin Li1, Pouya Parsa1, Shuang Zhou1, Xinheng Ding1, Chinmay Arvind1, Guanchu Wang2, Vladimir Braverman3, Ali Payani4, Yantao Zheng5, Zirui Liu1

1University of Minnesota    2University of North Carolina at Charlotte    3Johns Hopkins University    4Cisco    5Adobe

Preprint  ·  Corresponding author: zrliu@umn.edu

DynamicMem data pipeline: multi-timescale user profile construction, intent-conditioned event chain generation, and state-consistent multi-app log generation.
DynamicMem's data pipeline: it constructs an evolving, multi-timescale user profile, generates intent-conditioned event chains, and grounds them into state-consistent multi-app interaction logs.
15
months per user
2.2M
tokens / user (avg)
1,772
grounded events / user
16
applications
5
quarterly checkpoints

Abstract

LLM agents increasingly act as personal assistants that must remember a user's profile over months — who they are (attributes), what they routinely do (habits), and what they prefer (preferences) — and keep it updated as jobs, routines, and tastes drift. Existing benchmarks evaluate this "memory" through short, simplified interactions, missing three core properties of real long-horizon behavior: a user's profile is heterogeneous (attributes, habits, and preferences evolve on different timelines); its changes are causally driven by external context such as life events; and its evidence is scattered across many small actions in different apps rather than stated explicitly.

We introduce DynamicMem, a synthetic benchmark that constructs 15 months of activity for each user — the kind of long-term, multi-app data that real users' privacy keeps out of reach. It provides user-consistent trajectories averaging 2.2M tokens and 1,772 grounded events per user across 16 applications, with profiles that evolve under seasons and life events, and every recorded action reflecting the user's profile at that moment. We evaluate at five quarterly checkpoints to track how systems scale as history grows. Benchmarking five representative memory systems exposes problems a single accuracy score hides — detailed below.

How the Benchmark Works

A quick tour of the setting before the results.

Each user has a 15-month trajectory of activity across 16 apps. We place five quarterly checkpoints (C1–C5); at each one a memory system sees only the app logs up to that date and must answer about the user's state at that moment. Accessible history therefore grows from ~3 months at C1 to ~15 months at C5 — so later checkpoints reveal how a system scales as evidence piles up, and whether it updates facts that have changed.

One user's profile evolving across five quarterly checkpoints, with world-background drivers and typed state changes.
One user across the five quarterly checkpoints. A world background of seasons and life events (top) drives typed changes to the profile (bottom) — acquiring a weekly site-audit habit, adding an EPA-compliance attribute, shifting a learning preference. Every recorded action is consistent with the profile at that moment.

The profile decomposes into three state families

Attributes

Facts and possessions — what the user has or is. Change discretely on life events.

e.g. primary vehicle: a 2022 BMW 530e; home city: Pittsburgh.

Habits

Recurring routines — what the user regularly does. Adapt to context and may revert.

e.g. a Saturday-morning grocery run; a weekday 6 km jog.

Preferences

Comparative inclinations — what the user tends to choose. Drift gradually, rarely stated outright.

e.g. a leaning toward healthier food; quiet, self-paced learning.

Evidence is scattered across apps and time

Nothing in the profile is stated outright. A single user intent unfolds as a causally linked event chain across several apps over months, so the evidence for any one fact is fragmented — a memory system has to piece it back together.

A user intent unfolding as a causally linked event chain across Gmail, LinkedIn, and WhatsApp over two months.
One intent — obtaining and acting on an EPA VOC certification — spread across Gmail, LinkedIn, and WhatsApp over two months.

Two tasks: recover the state, then act on it

State Completion — recover the state

Fill the blank fields of the user's current profile in a fixed schema, condensing many time-scattered logs.

Personalized Service — act on the state

Handle a request whose scenario gives only time and place — the user-state needed is deliberately withheld.

e.g. "Sunday 10:15" → draft a reminder for the user's weekend grocery run at its usual time and place.

State Completion example: condensing scattered Fitbit walking logs into a fixed habit schema at the Q4 2023 checkpoint.
State Completion example: condense many time-scattered logs (left) into the blank fields of a fixed schema (right) — here, the user's current weekend-walk habit at the Q4 2023 checkpoint.

Both tasks are scored field by field by an LLM judge on a Core axis (is the central fact right?) and a Detail axis (are the specifics preserved?), combined as s = 0.8·Core + 0.2·(Detail/2).

Key Findings

Benchmarking Oracle plus five memory systems — RAG, HippoRAG2, A-Mem, MemoryOS, and SimpleMem — at the five quarterly checkpoints C1–C5.

Finding 1

State Completion degrades over the horizon — but Personalized Service does not.

From C1 to C5, State Completion declines for every system (−4.4 RAG, −6.6 HippoRAG2, −8.9 A-Mem, −14.2 MemoryOS, −16.7 SimpleMem). Yet Personalized Service shows no such decay — it is higher at C5 than C1 for four of the five systems. Since both tasks read the same memory, the bottleneck is recovering state, not acting on it.

State Completion and Personalized Service scores across the five quarterly checkpoints C1–C5 for all systems.
(a) State Completion and (b) Personalized Service across the five quarterly checkpoints C1–C5, for all systems. Markers are annotated with their change relative to C1.
Finding 2

The decline is mainly preferences; habits stay stable and attributes drop early then flatten.

Breaking State Completion down by state family, preferences account for most of the decline (e.g. −26.5 for A-Mem from C1 to C5) and degrade steadily — they are implicit, inferred from many scattered logs that later evidence blurs. Habits stay roughly flat (recurring behavior adds redundant evidence), and attributes drop early then hold.

State Completion across checkpoints C1–C5 decomposed by state family: preferences, habits, attributes.
State Completion across C1–C5, decomposed by state family: (a) Preferences, (b) Habits, (c) Attributes, for all systems. Markers are annotated with their change relative to C1.
Finding 3

Retention and update are separate failure modes — no architecture wins both.

A falling score can mean forgetting a long-standing fact (retention) or failing to adopt a changed one (update) — distinct failures an aggregate curve hides. The same mechanism that lets a system hold a stable fact often blocks it from adopting a changed one, and the trade-off differs by family: append-only stores (RAG, HippoRAG2) adopt changes fast but forget; consolidating stores (A-Mem, SimpleMem, MemoryOS) retain but blur old and new.

State Completion split into long-range retention (top row) and update (bottom row), by family.
State Completion split into long-range retention (top row) and update (bottom row), by family. Retention re-asks about facts unchanged since C1 (a forgetting curve as the lookback grows); update scores facts whose gold value just changed. Neither regime is defined at C1, so both rows start at C2.

Error Analysis

Every failure is labelled by what the memory system delivered to the answer model — isolating memory failures from answer-model behaviour.

Failure-type distribution per system on State Completion and Personalized Service.
Failure-type distribution per system on (a) State Completion and (b) Personalized Service. Each bar is 300 sampled failures (score < 1), classified into Irrelevant, Identity miss, Detail miss, Conflated, or All clear (evidence complete but the answer still wrong — i.e. an answer-model fault).
Finding 1

Structured memory shifts failures from detail to identity.

The four systems that build structure beyond raw chunks (A-Mem, HippoRAG2, MemoryOS, SimpleMem) all concentrate at 35–36% Identity Miss with Detail Miss falling to 28–33%; RAG shows the opposite balance (36% Detail Miss). Merging and summarizing preserve surrounding detail but abstract away the named anchor — and the identity-miss rate sits within a point of 35% across four very different mechanisms, suggesting it is a systemic challenge of structured memory.

Finding 2

Hierarchical summarization fails earlier in the pipeline.

MemoryOS shows the highest Irrelevant Evidence rate on both tasks (29% SC vs 23–24%; 40% PS vs 35–37%); combined with its Identity Miss, roughly two-thirds of its failures involve evidence that never pins down the user's state. Compacting many logs into a multi-level summary surfaces an abstracted view rather than the specific log a precise state query needs.

Finding 3

Improving memory is far more leverageable than improving the answer model.

All Clear — evidence complete and unambiguous, yet the prediction still wrong — stays at just 2–7% across all ten (system, task) combinations. In other words, more than 93% of failures are bounded by what the memory system could deliver, not by what the answer model did with it. The headroom lies in memory itself.

BibTeX

@article{xie2026dynamicmem,
  title   = {DynamicMem: A Long-Horizon Memory Benchmark in Real-World Settings},
  author  = {Xie, Wenya and Zhou, Shengming and Li, Zelin and Parsa, Pouya and
             Zhou, Shuang and Ding, Xinheng and Arvind, Chinmay and Wang, Guanchu and
             Braverman, Vladimir and Payani, Ali and Zheng, Yantao and Liu, Zirui},
  journal = {arXiv preprint arXiv:XXXX.XXXXX},
  year    = {2026}
}