ON-DEVICE MEMORY SURVEY PREPRINT · 2026

On-Device Agent Memory

Forms, Lifecycle Management,
and Personalization

Wenqi Hu1· Yuxiang Ren1,†· Ran Zhou1·Qijia Zhuang1·Zihan Shen1·Yanyin Gou1·Zehui Kong1· Fanyu Lin2·Jiale Chang3·Zhen Zhang1·Jiaheng Liu1·Haopeng Zhang4·Jiawei Zhang5·Philip S. Yu6
1Nanjing University2Peking University3Nanjing Agricultural University4UNC Charlotte5University of California, Davis6University of Illinois Chicago

† Corresponding author

Updated October 18, 2026 · 56-page manuscript · Original figures & open reading resources

5Memory forms
4Lifecycle operations
4Platform families
215Cited works
32Benchmark entries

From personal experience
to useful memory.

What should an agent remember, how should that memory change, and when should it shape an action? The survey connects these questions across five perspectives.

View the complete taxonomy from the paper +
Explore the figure
Figure 1-2. The survey taxonomy and representative works, reproduced from the latest manuscript.

SECTION 1 · INTRODUCTION

Memory, under everyday constraints.

Personal devices bring together continuous multimodal experience, bounded resources, and a user's long-term history. Useful agent memory must connect all three.

Explore the figure
Figure 1-1. Typical cloud and personal-device deployments across inputs, resource budgets, and personalization. Click any figure to enlarge ↗

This survey connects memory forms and lifecycle management to personalization, then examines deployment evidence, evaluation, and the challenges of sustained use on personal devices.

SECTION 2 · FOUNDATIONS

What is on-device agent memory?

Read Section 2 ↗

Memory as an evolving state

Agent memory keeps information from past observations, actions, and feedback available to the current decision process.

HistoryMemoryAction
Episodic
Specific experiences and the context in which they occurred.
Semantic
Facts and relationships usable beyond a single episode.
Procedural
Reusable skills and ways of carrying out an activity.

Short-term and long-term memory describe retention; working memory describes the active use of information in an ongoing task.

Local memory, within explicit limits

On-device agent memory is retained and processed on a personal device under resource budgets and data-use rules.

Peak runtime memory covers the agent as a whole: memory processing and model inference draw on the same available capacity.

Data-use rules

Specify what may be retained, for how long, and for which tasks. Later access, disclosure, correction, and deletion remain subject to user authorization.

Five ways to retain experience.

A form defines the units in which memory is stored and reused. These forms can coexist: a personal fact, a reusable skill, and a repeated context need not share the same representation.

Explore the figure
Figure 3-1. Five memory forms and the subcategories used to organize the review.
EXPLICIT & INSPECTABLE

Textual memory

Selected experience is retained as natural language or structured text, making personal information available for direct inspection and revision.

CATEGORIES IN THE SURVEY
  • Event records
  • Hierarchical summaries
  • Task-state records
  • User profiles
  • Procedural experience

Memory is a process.

A memory store needs more than a write path. Its contents must enter tasks at the right time, incorporate new evidence, and leave when they are no longer useful or permitted.

01

Formation

Decide what enters memory, in which units, and with what provenance.

Admission · Units · Provenance
02

Retrieval

Decide when to access memory, where to look, and how much evidence to return.

Triggering · Routing · Reduction
03

Evolution

Consolidate jointly valid evidence and update facts or procedures as conditions change.

Consolidation · Updating
04

Forgetting

Control continued residence and influence through selection, deletion, and unlearning.

Selective · Directed · Unlearning
Explore the figure
Figure 4-1. Filled arrows indicate memory changes; outlined arrows indicate information flow.

Remembering, for a particular person.

Personal histories help an agent interpret preferences, resolve references, and anticipate needs. Privacy safeguards constrain how that history is retained and used throughout the system.

Explore the figure
Figure 5-1. Roles of personal memory and the privacy safeguards that constrain them.
PREFERENCES

What do I tend to choose?

Learn from explicit feedback and repeated behavior, while allowing preferences to change with context and time.

PERSONAL REFERENCES

What does this mean to me?

Link a request to the people, places, possessions, and experiences in a particular user's world.

PROACTIVE ASSISTANCE

Is this a useful moment to help?

Use recurring intent and present conditions to judge whether an intervention is timely and welcome.

Different devices. Different memory demands.

Four platform families connect memory design to the conditions in which agents operate. Device setting and execution location are examined separately.

PERSONAL COMPUTING

Phones & computers

Transient interface state, cross-application history, and local files support tasks that outlive a single screen.

FOCAL · MOBIMEM · PalmClaw
HEAD-WORN

Smart glasses & XR

First-person sensing requires selective capture; immersive sessions require spatial and interaction continuity.

EgoLife · EgoTrigger · ClickAIXR
PERSONAL SENSING

Watches, health & assistance

Intermittent physiological, motion, and acoustic signals become event records and longer-term personal context.

SAMoSA · PHIA · MemPal
EMBODIED

Robots & simulated agents

Maps, object states, and reusable skills inform navigation and manipulation within perception–action loops.

VLONS · EfficientNav · Voyager

↳ A study's inclusion in a device context does not by itself establish end-to-end on-device execution. Section 6 distinguishes local components, remote computation, physical experiments, and simulation.

Measure what memory
makes possible.

Evaluate information quality and useful behavior alongside the resources needed to sustain them.

01

Memory accuracy

Recall, evidence access, and reasoning over history.

02

Lifecycle effectiveness

Formation, updating, retention, and forgetting.

03

Personalization quality

User-specific interpretation and task outcomes.

04

On-device runtime cost

Memory footprint, latency, energy, and maintenance.

Benchmark explorer 32

General and device-context entries from Table 7-1.

Read Section 7 ↗

32 benchmark entries

BenchmarkScenarioEvaluation focusResources

Links reproduce the latest manuscript. “Paper only” means Table 7-1 does not specify a separate resource link.

What remains open?

Sustained use on personal devices raises questions that extend beyond a single successful retrieval.

01

Resource-aware lifecycle management

Coordinate retrieval and background maintenance as memory, workloads, and available resources change.

02

Reliable memory evolution

Preserve provenance and meaning through successive summaries, updates, and derived records.

03

Continual personalization

Carry uncertainty forward and revise personal preferences as circumstances and feedback evolve.

04

Privacy protection in memory use

Carry restrictions through derivation and reassess permitted use as tasks and purposes change.

05

Verifiable forgetting

Check removal from stored records and learned influence while preserving unrelated capabilities.

06

Long-term evaluation on real devices

Connect controlled histories to sustained interaction, accumulated costs, and the user's management burden.

Follow the evidence.

Search all 215 distinct works cited in the latest manuscript. Topic labels indicate where a work is cited in the survey.

Download references.bib ↓

215 cited works

Cite this survey.

Bibliographic information for the current preprint.

BibTeX
Download .bib

100%
Original ↗

Use the zoom buttons, then scroll in any direction to explore. The plus and minus keys also change the zoom; zero fits the width.

Benchmark explorer