AuthorsDeepak Babu Piskala
TL;DR
MAPLE decomposes agent memory into three sub-agents for Memory, Learning, and Personalization, yielding a 14.6% personalization score gain and +30pp trait incorporation on MAPLE-Personas.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM assistants often store feedback but fail to adapt, as shown by MAPLE-Personas where stateless baselines reach only a 4.17 personalization score and 45% trait incorporation.
This conflation causes systems like enterprise AI assistants to ignore user preferences, producing identical responses for Sarah and Marcus and leading to frustration and lost value despite rich interaction histories.
HOW IT WORKS
MAPLE introduces three sub-agents: Memory for storage infrastructure, Learning for intelligence extraction, and Personalization for real-time adaptation, all coordinated by an Agent orchestrator.
Conceptually, MAPLE treats Memory as a warehouse, Learning as a curator, and Personalization as a shopper, mirroring RAM vs disk or hippocampus vs cortex in human cognition.
This decomposition lets MAPLE run Learning asynchronously while Memory and Personalization operate in the request path, enabling continual adaptation that a plain context window or monolithic “memory” cannot provide.
DIAGRAM
This diagram shows how MAPLE handles a user query in real time while asynchronously updating Memory through the Learning loop.
DIAGRAM
This diagram shows how MAPLE-Personas generates personas, runs Baseline vs MAPLE, and uses an LLM judge to score personalization and trait incorporation.
PROCESS
Request Time Personalization Flow
MAPLE receives a query at the Agent, calls Personalization, which retrieves user preferences from Memory and assembles a tailored context for generation.
Selective Retrieval from Memory
The Memory sub-agent queries token level and hierarchical stores, returning only traits relevant to the current question, such as technical depth for Sarah.
Personalization Mechanism
The Personalization sub-agent composes instructions and user context, telling the LLM to adapt style and content based on the explicit user model U𝑢.
Asynchronous Learning Loop
After sessions, Learning analyzes episodic histories, infers preferences and patterns, and writes structured insights back into Memory for future personalization.
KEY CONTRIBUTIONS
MAPLE Sub-Agent Decomposition
MAPLE decomposes adaptation into Memory, Learning, and Personalization sub-agents, each with its own tools and LLM instance, enabling modular optimization and failure isolation.
MAPLE-Personas Benchmark
MAPLE introduces MAPLE-Personas with 150 personas and 10 turn dialogues, measuring proactive personalization via judge scores and trait incorporation across Baseline and MAPLE systems.
Quantified Personalization Gains
MAPLE achieves a mean judge score of 4.78 vs 4.17 for the Baseline and raises trait incorporation from 45% to 75%, with Cohen’s d = 0.95 and 𝑝 < 0.01.
RESULTS
Judge Score (1–5)
4.78 score
+0.61 over Baseline
Trait Incorporation
75%
+30pp vs Baseline
Perfect Personalization
88%
vs 15% Baseline perfect scores
Effect Size Cohen d
0.95
large effect with Welch t = 6.05, p < 0.001
On the MAPLE-Personas benchmark, which tests proactive personalization across held out turns, MAPLE’s 4.78 judge score and 75% trait incorporation show that MAPLE’s sub-agent architecture materially improves user specific adaptation over the Baseline.
BENCHMARK
On the MAPLE-Personas benchmark, which tests proactive personalization across held out turns, MAPLE’s 4.78 judge score and 75% trait incorporation show that MAPLE’s sub-agent architecture materially improves user specific adaptation over the Baseline.
BENCHMARK
Judge personalization score and trait incorporation rate comparing MAPLE to the Baseline configuration.
KEY INSIGHT
MAPLE, using a simple filesystem based Memory with JSON records instead of vector databases, still boosts mean judge scores from 4.17 to 4.78 and trait incorporation from 45% to 75%.
This is surprising because many practitioners assume sophisticated retrieval infrastructure is necessary, but MAPLE shows that architectural decomposition of Memory, Learning, and Personalization drives the gains instead.
WHY IT MATTERS
MAPLE unlocks agents that genuinely learn from interaction histories via symbolic Learning and explicit user models, then personalize in real time through a dedicated Personalization sub-agent.
Builders can now plug MAPLE’s sub-agent architecture into existing stacks, swap storage backends, and still get measurable personalization improvements without fine tuning or opaque parametric memory updates.
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…