Hermes memory setup: Mnemosyne vs Hindsight

This tutorial compares Mnemosyne and Hindsight as memory providers for Hermes Agent. The speaker separates operational memory, such as working preferences, from world knowledge kept in an Obsidian vault or LLM Wiki. He explains how Hermes injects its built-in memory files into new sessions, while external providers retrieve relevant information when needed.

The Mnemosyne walkthrough covers a local SQLite memory layer that the speaker describes as requiring no LLM. He installs it in a separate pipx environment because Hermes updates rebuild the managed environment. His setup also needs a link to the Hermes plugins folder and the memory setup command. He selects global recall across sessions, disables legacy memory to avoid duplication, and tests whether a new chat remembers his profile.

For Hindsight, he creates a separate Hermes profile and runs a self-hosted server in Docker. The demonstrated configuration uses Ollama with GPT-OSS 20B, described as a 13 GB model. The model must support tool calling, and Ollama and Docker Desktop need to be running. He connects Hermes through local external mode, then tests retention and recall and explores the memory dashboard.

Hindsight's reflection step uses an LLM to synthesize stored information. The speaker reports an issue during that example but finds the response useful. His comparison favors Mnemosyne for simplicity on one machine and Hindsight for deeper memory processing.