Paper Tests File-System Memory for LLM Agents: Halves Retrieval Cost, Degrades Over Time
dair_ai · x · 2026-07-31
A new paper evaluates whether file-system-based long-term memory (a folder of markdown files managed via standard file tools) is effective for LLM agents.
The study tested two previously unchecked assumptions: whether an agent can keep a growing store organized as memories accumulate, conflict, and go stale, and whether this organization pays for itself.
Key Findings:
- Organized stores roughly halve retrieval costs when dealing with large amounts of material.
- No tested agent successfully converted this organization into better final answers.
- In growth scenarios, the memory store degraded for every management agent except the strongest one.
- Simply changing the toolset reshapes the memory store as profoundly as swapping the underlying model.
More from coding & agent
- Marble (YC S26): AI Agents to Automate Restaurant Back of House — ycombinator · 2026-07-31
- Microsoft Releases Echoverse: Deep Evolving Environments for Evaluating Computer-Use Agents — Scobleizer · 2026-07-31
- Salesforce Unveils Metadata API Skill to Cut Deploy Errors — msrivastav13 · 2026-07-31
- Unexpected Prompt Injection: When Generating Presentations with Examples — tristanbob · 2026-07-31
- SpecFirst Framework: Agents Write Specs First, Boosting Code Synthesis by 21% — centre-for-swe · 2026-07-31
- VP PM's Deep Dive: 4 Real Pain Points of Coding with LLMs and the Human Edge — Greedy_Rise_6567 · 2026-07-31