News · 2026-07-31

Letting an agent organise its own memory halved retrieval cost and improved no answers

A team spanning the University of Illinois, UC San Diego and Adobe published the first systematic study of the memory design that deployed AI agents actually use - a directory of markdown files the agent maintains itself - and found that organising it reliably halves retrieval cost while improving nothing about the answers. In their growth study, organisation eroded for every management agent except the strongest, and no agent they measured converted a tidier store into better responses.

Key facts

• Organised memory stores roughly halve retrieval cost where the material is large.
• No agent measured converted organisation itself into better answers; organisation degraded as stores grew for all but the strongest management agent.
• Published July 29, 2026, by Sizhe Zhou, Sheldon Yu and colleagues at Illinois, UC San Diego and Adobe.
• Primary source: Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability.

If you have used a coding agent recently, you have used this design without being told. The agent keeps notes in markdown files, reads them back at the start of a session, and rewrites them as it learns. It is the most widely deployed form of agent memory there is, and as the authors point out, research had skipped past it entirely in favour of bespoke memory architectures - leaving two assumptions untested. First, that an agent can keep a growing store organised as memories accumulate, conflict and go stale. Second, that this organisation pays.

The study splits the problem into three roles around one shared filesystem: a management agent that integrates and reorganises incoming content, a search agent that answers queries with citations, and an execution agent whose task trajectories get distilled into reusable skills. Then it varies almost everything - memory shape, the volume of incoming material, the tool set the agent is given, and the strength of the management and search agents - while tracking answer quality, cost and what the authors call store health.

The result on cost is clean and positive: "What organization reliably buys is search economy: organized stores roughly halve retrieval cost where material is large." That is a real win. Retrieving from a well-structured directory takes fewer reads and fewer tokens than grepping a pile.

The result on quality is the one that matters more, and it is negative. "Today's agents, however, fall short of the default's promise: in our growth study, organization erodes for all but the strongest management agent, and no agent we measure converts organization itself into better answers." Strength of the search agent mattered; tidiness of the store did not. That inverts the intuition most builders work from - the belief that if you can just get the agent to keep good notes, the answers will follow.

The analogy is a filing cabinet that gets messier as more paper arrives. Anyone can organise ten documents. Organising a thousand while a hundred more arrive daily, several contradicting what is already filed, is a different job - and it is the job that degraded in this study. Meanwhile the person who is good at finding things did fine in either cabinet.

The third finding is the least expected and the most actionable: "the model is not the only lever over a store's shape: changing the tool set alone reshapes the store as strongly as swapping the model." Give an agent one set of file operations and it produces a few large documents; give it another and it produces many small ones. That is a design decision most teams make by accident, in an afternoon, while wiring up tools - and it turns out to be as consequential as choosing which model to run.

The honest caveat is scope: these are long-conversation benchmarks and embodied tasks over specific growth horizons, not a general law about memory. A store that erodes over this study's timescale might stabilise over another, and a better management agent might yet convert structure into quality - the strongest one tested already resisted the erosion. What the paper establishes is narrower and still useful: the filesystem default has been an assumption, and measured against its own promise, half of that promise is currently unearned.

Primary source, verified: read the paper → (arXiv 2607.26637)

Ground Truth. (2026, July 31). Letting an agent organise its own memory halved retrieval cost and improved no answers. Ground Truth. https://groundtruth.day/news/organising-an-agents-memory-halves-search-cost-but-does-not-improve-answers.html

@misc{groundtruth:organising-an-agents-memory-halves-search-cost-but-does-not-improve-answers,
  title  = {Letting an agent organise its own memory halved retrieval cost and improved no answers},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/organising-an-agents-memory-halves-search-cost-but-does-not-improve-answers.html}
}

Topics: agent-memory · agents · evaluation · research · negative-result