Why do memory systems perform so well in narrow domains?
In narrow, repetitive domains, the patterns an agent needs to remember are predictable and bounded. A medical QA system with a cache-and-prune memory bank achieved 82.98% on USMLE Step 1 and 86.24% on Step 2, outperforming GPT-4's 80.67% and 81.67% respectively [1]. The memory system didn't just store facts—it reused reasoning workflows specific to medical diagnosis. Similarly, in web navigation across 1,000+ tasks from 200+ domains (travel, shopping, social media), a method called Agent Workflow Memory (AWM) improved success rates by 24.6% on Mind2Web and 51.1% on WebArena, while also reducing the number of steps needed [2]. These are large, practical gains because the tasks, though varied, share underlying action patterns that memory can capture.
The reason is structural: narrow domains have a limited set of valid actions and outcomes. A GUI agent for Android phones used a structured knowledge graph to compress multi-step routines, outperforming UI-TARS-7B by up to 19.6% on AndroidWorld while cutting token costs by 6x relative to GPT-4o [9]. The memory system essentially learned 'recipes' for common tasks, which is far easier when the recipe book is finite. In image generation workflows, a skill-evolution system (COMFYCLAW) distilled past trajectories into reusable skills, achieving the best average evaluation score across all six agent configurations tested [4]. Human annotators also preferred the version with skill evolution, confirming that in a narrow creative domain, memory directly improves reliability.
What happens when you take memory systems out of their comfort zone?
When memory systems are tested across diverse, unpredictable scenarios, most of them stumble. A 2026 study revisited eight memory systems plus a simple agentic harness across five very different scenarios: single-turn QA, multi-session chat, agentic-trajectory QA, memory stress tests, and long-horizon agentic tasks [8]. The result was striking: no single memory design dominated. The best cross-scenario performer was not a sophisticated memory architecture but a simple harness that gave the agent active control over its own flat text-file storage via tool calls. This suggests that in general-purpose workflows, rigid memory structures can become a liability—they optimize for patterns that don't hold.
Another study found that existing memory systems are typically 'tuned to a single scenario' and show 'little evidence that they generalize across the heterogeneous trajectories agents encounter in deployment' [8]. This is a direct admission from the research community that narrow-domain success hasn't yet translated to broad utility. Even the most advanced systems, like the schema-constrained generative memory (SCG-MEM), which formally guarantees no structural hallucinations, was only evaluated on a single benchmark (LoCoMo) [6]. The gap between best-case and typical-case evidence is real: a system that crushes one benchmark may fail on the next.
Are researchers finding ways to make memory work across domains?
Yes, several new approaches aim to bridge the narrow-to-general gap, and early results are promising. The General Agentic Memory (GAM) framework uses a 'just-in-time' principle: it keeps only lightweight memory offline, then at runtime a 'Researcher' component retrieves and integrates detailed information from a universal page-store [5]. This design achieved 'substantial improvement' across various memory-grounded tasks, suggesting that separating what you store from how you retrieve at runtime may be key to generality.
Another promising direction is cross-paradigm alignment. MemAdapter can unify explicit, parametric, and latent memory paradigms within a single system, completing cross-paradigm alignment in just 13 minutes on a single GPU [3]. It outperformed five strong agent memory systems across three paradigms and multiple model scales. This matters because a general-purpose agent will likely need to mix memory types—some facts, some skills, some patterns—and MemAdapter shows it's possible to fuse them efficiently.
Perhaps the most practical insight comes from the 'decoupling before aggregation' principle behind xMemory [7]. Instead of retrieving whole chunks of past interaction, xMemory first isolates reusable facts and distinguishing details, then organizes them hierarchically. It showed consistent gains in answer quality and inference token efficiency across both open-source and closed-source LLMs on two benchmarks. This approach—breaking memory into atomic, reusable pieces—may be the design pattern that finally makes memory systems useful in general workflows, because it doesn't assume the task type in advance.
About These Sources
This answer is built on 9 studies (1 peer-reviewed, 8 preprints) — published from 2024 to 2026, 9 from 2024 or later, 1 in Q1 journals — selected as the most relevant from 15 studies that passed quality screening, drawn from 48 papers retrieved from a database of over 500 million.
Sources used in this answer
Agentic memory-augmented retrieval and evidence grounding for medical question-answering tasks
A memory-augmented agentic system outperformed GPT-4 on USMLE Step 1 (82.98% vs. 80.67%) and Step 2 (86.24% vs. 81.67%) across five medical QA benchmarks, showing narrow-domain strength.
Agent Workflow Memory
Agent Workflow Memory (AWM) improved web navigation success rates by 24.6% on Mind2Web and 51.1% on WebArena across 1,000+ tasks from 200+ domains, while reducing steps.
MemAdapter: Fast Alignment across Agent Memory Paradigms via Generative Subgraph Retrieval
MemAdapter unified three memory paradigms (explicit, parametric, latent) in 13 minutes on a single GPU, outperforming five strong baselines across multiple model scales.
COMFYCLAW: Self-Evolving Skill Harnesses for Image Generation Workflows
COMFYCLAW, a skill-evolution system for image generation workflows, achieved the best average evaluation score across all six agent configurations, with human annotators preferring it.
General Agentic Memory Via Deep Research
General Agentic Memory (GAM) uses a just-in-time principle with a Memorizer and Researcher, achieving 'substantial improvement' on various memory-grounded tasks.
To Know is to Construct: Schema-Constrained Generation for Agent Memory
SCG-MEM, a schema-constrained generative memory, provides formal guarantees against structural hallucinations and improved performance on the LoCoMo benchmark.
Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation
xMemory, based on 'decoupling before aggregation,' showed consistent gains in answer quality and inference token efficiency across open-source and closed-source LLMs on two benchmarks.
Exploring Cross-Scenario Generality of Agentic Memory Systems: Diagnostics and a Strong Baseline
Across eight memory systems tested on five diverse scenarios, no single design dominated; the best cross-scenario performer was a simple harness giving the agent active control over storage.
Executable Agentic Memory for GUI Agent
Executable Agentic Memory (EAM) for GUI agents outperformed UI-TARS-7B by up to 19.6% on AndroidWorld while reducing token costs 6x relative to GPT-4o, with 2.8s average latency.