Paper: Filesystem-Based Memory for LLM Agents — Organization, Evolution, and Sustainability
Sizhe Zhou · hf · 2026-07-31
Deployed LLM agents increasingly use a directory tree of markdown files as long-term memory. This study presents the first systematic exploration of this filesystem-based memory, testing if agents can sustain organization as memories accumulate and whether it pays off.
The setting is formalized into three roles: a management agent (integrates content), a search agent (answers with citations), and an execution agent (supplies trajectories distilled into skills). Key findings from benchmarks include:
- Search Economy: Organized stores roughly halve retrieval costs when material is large.
- Organization Erosion: For all but the strongest management agents, organization degrades over time, and no agent converts organization itself into better answers.
- Toolset Impact: Changing the tool set alone reshapes the memory store as strongly as swapping the underlying model.
Related event: ICML Paper Tests File-System Memory for LLM Agents(2 posts)→
More from coding & agent
- Implementing OAuth 2.1 for Remote MCP Servers: A Developer's War Story — Ranorkk · 2026-07-31
- Hugging Face Expert Shares: Training Agents by Distilling Code Traces — mervenoyann · 2026-07-31
- Kaggle Launches Kaggriculture Simulation Contest for Autonomous Agents — JFPuget · 2026-07-31
- Building Reliable AI Agents with Google ADK 2.0 and A2UI — SucceededMind · 2026-07-31
- Blockstream Uses Frontier LLMs to Uncover Crypto Hardware Vulnerabilities — RSync25 · 2026-07-31
- Integrating DeepSeek-V4-Flash into Codex: Costs 89x Less Than Opus — teortaxesTex · 2026-07-31