Hindsight setup: shared memory for Hermes and Codex

This tutorial sets up Hindsight as a shared memory service for Hermes and Codex. The speaker demonstrates both directions: Hermes recalls a temporary project color saved through Codex, and Codex retrieves an animal preference saved through Hermes.

The self-hosted setup uses Docker Compose, with Docker, Python and both agents installed beforehand. An environment file holds separate keys for agent API access and the control plane. Hindsight also needs an LLM for memory processing. This walkthrough uses a dedicated Docker volume for Codex authentication and local embeddings for retrieval; the speaker refers to an earlier Ollama setup as a local model option. Hosting the memory service locally does not make the demonstrated model processing offline.

The configuration covers retention, recall, reflection, entity labels and sensitive-content redaction. The speaker uses one global bank for related projects, explaining that banks form recall boundaries while tags allow filtering within a bank. Automatic project tagging can make mistakes.

Hindsight v0.9.0 introduces the coding agent plugin used to connect Codex through hooks and MCP. The tutorial also covers backups, upgrades and knowledge pages generated from stored memories. The speaker treats these pages as agent-maintained working context and keeps human-reviewed documentation in Obsidian as a separate trust layer.