You installed an AI coding assistant. You ran a few local agents. Now your storage is disappearing and the cause is not obvious — because AI agents do not just use model files. They create storage of their own.
What agents actually store
A local AI agent workflow typically involves several layers of storage, each managed by a different tool:
- The model backend — usually Ollama or a similar runner. This stores the actual model weights. The largest files by far.
- Agent workspace and context — many agent frameworks keep local project memory, conversation history, or task state. These grow with every session.
- Embeddings and indexes — if an agent indexes your codebase or documents for retrieval, it generates embedding vectors stored locally. These can grow large for big projects.
- Logs and traces — agent frameworks often log every action, tool call, and response for debugging. Those logs accumulate silently.
- Caches — intermediate results, API response caches, or downloaded tool dependencies.
Each layer is managed by a different tool and stored in a different location. No single overview exists by default.
Want to skip the hidden-folder hunt?
LLM Cleaner scans your Mac for local AI models, caches, indexes, and project memory — then shows what you can review, reveal, export, or safely move to Trash.
The tools involved
AI agent workflows in 2025 and 2026 commonly involve some combination of:
- Ollama — running models locally, storing weights in hidden blob directories
- Claude Code — maintaining project memory and context files per workspace
- Cursor — workspace indexes and AI context per project
- Windsurf / Codeium — AI suggestion caches and workspace state
- OpenCode, Codex, and similar tools — model caches and conversation history
Each of these stores files in a different place. None of them communicate with each other about shared storage. And none of them offer a clear disk-usage overview.
Why standard disk cleaners miss this
A generic disk cleaner looks for large files. It will find model blobs if they are large enough. But it will not understand what those blobs are, whether they are safe to remove, or which agent workspace they support.
Deleting a model blob that an active agent workflow depends on does not produce an obvious error. It produces a subtle failure — a degraded response, a missing context, an agent that no longer works — that you may not trace back to the cleanup for days.
Local AI storage is messy. Cleanup should not be.
Find Ollama, LM Studio, Hugging Face, Cursor, Windsurf, Claude Code, Whisper, and loose model files in one place.
The review-before-delete principle
With model files and agent storage, the right approach is always: see everything first, understand what each item is, then decide what to remove — not the other way around.
That is especially true for agent caches and project memory, where some files are genuinely valuable context and others are stale logs from experiments you finished months ago.
Not sure what is safe to delete?
LLM Cleaner separates models, rebuildable caches, and project memory so you do not treat everything like junk.