AuthorsJinghao Luo, Yuchen Tian, Chuxue Cao et al.
TL;DR
From Storage to Experience uses cross trajectory abstraction to unify LLM agent memory evolution into three stages without a shared benchmark taxonomy.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM based agents suffer from statelessness that “precludes learning from prior interactions, often resulting in recurring reasoning errors.” This blocks long horizon tasks.
Existing memory work “oscillates between operating system engineering and cognitive science,” leaving no coherent evolutionary perspective or technological synthesis for builders.
Without unified Storage, Reflection, and Experience stages, agents cannot achieve continual learning or robust operation in dynamic environments.
HOW IT WORKS
From Storage to Experience introduces three stages of memory mechanisms: Storage, Reflection, and Experience, with sub mechanisms like Linear, Vector, Structured storage and Introspection, Environment, Coordination reflection.
You can think of Storage as disk logs, Reflection as a debugging pass that cleans and compresses logs, and Experience as distilled playbooks or firmware updates guiding future behavior.
This cross trajectory abstraction lets From Storage to Experience move beyond a plain context window, turning dense trajectories into reusable rules and skills that act as policy priors for unseen tasks.
DIAGRAM
This diagram shows how an LLM agent interacts with a dynamic environment while retrieving and updating memory across Storage, Reflection, and Experience.
DIAGRAM
This diagram shows the top down taxonomy of memory mechanisms and how they evolve from trajectory preservation to abstraction.
PROCESS
Storage
From Storage to Experience first records trajectories into Linear, Vector, or Structured storage, preserving observation action pairs with minimal transformation.
Reflection
From Storage to Experience applies Introspection, Environment, and Coordination reflection to transform each trajectory τi into refined memory units m′i using evaluation criteria ϕ.
Experience
From Storage to Experience groups similar trajectories Tbatch and runs cross trajectory abstraction to derive universal rules or skills K as Explicit, Implicit, or Hybrid experience.
Active Exploration
From Storage to Experience then uses these experiences to drive active exploration, collecting new trajectories that feed back into Storage, Reflection, and Experience for continual learning.
KEY CONTRIBUTIONS
Three stage memory mechanisms
From Storage to Experience formalizes LLM agent memory into Storage, Reflection, and Experience, distinguishing raw trajectory preservation, semantic refinement, and cross trajectory abstraction.
Evolutionary drivers and path
From Storage to Experience analyzes long term consistency, dynamic environments, and continual learning as drivers, then maps them onto Linear, Vector, Structured storage and reflection modes.
Transformative Experience stage
From Storage to Experience identifies Active Exploration and Cross Trajectory Abstraction as core mechanisms, highlighting Explicit, Implicit, and Hybrid experience for future LLM agents.
RESULTS
Stages defined
3 stages
+3 conceptual layers over prior flat memory surveys
Core drivers
3 drivers
long term consistency, dynamic environments, continual learning
Experience modes
3 modes
Explicit, Implicit, Hybrid experience categories
Key mechanisms
2 mechanisms
active exploration and cross trajectory abstraction
From Storage to Experience is a survey and does not report benchmark metrics. Instead, it structures the field into three stages, three drivers, three experience modes, and two transformative mechanisms, giving builders a quantitative sense of the taxonomy’s scope.
BENCHMARK
From Storage to Experience is a survey and does not report benchmark metrics. Instead, it structures the field into three stages, three drivers, three experience modes, and two transformative mechanisms, giving builders a quantitative sense of the taxonomy’s scope.
BENCHMARK
Conceptual count of primary mechanisms in each memory stage.
KEY INSIGHT
From Storage to Experience argues that “unrestricted expansion of memory is detrimental,” as dense storage can propagate errors and contaminate learning. This challenges the intuition that more memory is always better.
This is counterintuitive because many builders assume scaling Vector or Structured stores monotonically improves agents, but From Storage to Experience shows it can degrade reasoning and consistency over time.
The survey instead pushes toward strategic addition and deletion and high level Experience abstraction, breaking the assumption that raw trajectory accumulation is harmless.
WHY IT MATTERS
From Storage to Experience gives a clear scaffold for designing LLM agents that move from Storage to Reflection to Experience, instead of bolting on ad hoc memory modules.
Builders can now map concrete systems like vector databases, reflection loops, and skill libraries onto this framework, systematically adding Active Exploration and Cross Trajectory Abstraction rather than guessing at memory designs.
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…