Hermes Agent memory setup with Honcho and Obsidian

This tutorial explains four memory layers in Hermes Agent: built-in markdown files, conversation search, an external provider, and an Obsidian vault. The speaker presents them as complementary parts of an AI agent setup.

The walkthrough starts with memory.md and user.md in .hermes/memories, with a combined budget of about 1,300 tokens. Both enter the system prompt at session start. Memory writes update disk immediately, but the prompt keeps its starting snapshot until the next session to preserve prompt caching. The speaker also explains character caps, save/skip rules, duplicate detection and injection scanning. FTS5 search retrieves earlier conversations from state.db; the supplied metadata identifies automatic pruning and vacuuming at startup as a v0.11+ feature.

The provider comparison covers Mem0's fact extraction, Hindsight's knowledge graph and SuperMemory's separate memory containers. Only one provider can be active at a time, and switching does not migrate stored data. The Honcho demonstration uses cloud hosting and an API key through hermes memory setup. The speaker also describes a self-hosted option and discusses the Honcho changes in version 0.11.0.

The final example uses the bundled Obsidian skill to write linked markdown notes for HVAC provider research. The filesystem workflow can run on headless Linux without the Obsidian app or an MCP server. The research produces provider notes and equipment types, but does not find specific equipment models.