Hindsight setup: local agent memory with Gemini options

This tutorial introduces Hindsight as an open source memory system for a coding assistant. The presenter walks through a Python virtual environment and a sample project with a choice of local model execution or Gemini. The cloud option reads an API key from an environment file; the example code falls back to a local model if that key is missing.

The explanation covers four memory areas: persistent facts, task history, working assumptions and refreshed summaries. Retrieval combines semantic matching, keyword searches, connections between records and date filtering. A sample product pitch supplies the context for investor, senior engineer and customer support personas. The project saves their responses as a Markdown report and an HTML dashboard.

The presenter reports 94.6% benchmark accuracy with Gemini 3 Pro and 83.6% with an unnamed 20-billion-parameter open model. These are attributed results, not performance guarantees for the setup shown. The walkthrough runs its example through Gemini, so its outputs do not establish equivalent local results.

For users building an AI agent, the resource and cost distinctions matter. The speaker claims the Python library uses less than 400 MB of memory and downloads roughly 300 MB of weights on first launch. Searches are described as local, but saving memories triggers a language model call for each text chunk. Cloud-backed ingestion can therefore incur costs; the presenter recommends batching text.