Fix model cache location to use memory directory instead of current working directory

[nikomatsakis]

Fix Model Cache Location

Status: Planning

Current Understanding

The Ollama model caching currently stores downloaded models in the current working directory where the hippo server is run. This creates several problems:

1. Inconsistent storage - Models and memories are stored in different locations
2. Permission issues - Current directory might not be writable
3. Clutter - Model files appear in random project directories
4. Cleanup difficulty - Hard to find and manage cached models

The model cache should be stored in the same directory as the memories (typically ~/.hippo/) for consistency and proper organization.

Next Steps

• Locate where model caching directory is configured in the Rust code
• Update model cache path to use the same --memory-dir argument
• Test that models download to the correct location
• Update documentation if needed
• Consider migration strategy for existing cached models

Open Questions

• Should we automatically migrate existing cached models from old locations?
• What's the current default cache location behavior in ollama-rs?
• Do we need to create a models/ subdirectory within the memory directory?

Context

This issue was discovered while reviewing the automatic RAG-like discovery implementation (issue #9). The model caching behavior is inconsistent with our memory storage approach and can cause user confusion and system administration problems.

This is a bug fix that should be addressed before continuing with the RAG implementation to ensure clean, predictable file organization.