The paper presents the first systems characterization of for LLM agents. The authors introduce a taxonomy classifying memory systems along four axes (construction, storage, retrieval, mutability) and build a phase-aware profiling harness that attributes costs to construction, retrieval, and generation phases. They characterize ten representative systems across two benchmark suites, revealing that construction (the write path) dramatically dominates costs—especially in embedding and prefill—while design choices shift costs asymmetrically between write and read paths. This work is significant because accuracy-only benchmarks miss these system-level tradeoffs, which are critical for deploying stateful agents at scale; the paper derives ten system recommendations covering scheduling, amortization, freshness-latency tradeoffs, and fleet management.
First systematic characterization of agent memory systems across long-horizon tasks
Researchers propose a taxonomy, profiling harness, and benchmark evaluation revealing cost tradeoffs between write and read paths.
Top University
Yasmine Omri · Ziyu Gan · Zachary Broveak · Robin Geens · Zexue He · Alex Pentland · +3 more
Stanford University · Independent Researcher · MICAS, KU Leuven · Massachusetts Institute of Technology
Research Digest··2 min read
Thread:Memory Management for Agents
In one line
Construction, not generation, is the dominant cost in agent memory workloads.
Before you read this
Two ideas this paper builds onRead it as
The authors present the first systems-level characterization of LLM agent memory systems.
The authors developed a system-oriented taxonomy to classify systems along four axes.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.