Hindsight: agent memory and local Docker setup overview

Hindsight is Vectorize's open source memory system for an AI agent. This repository walkthrough explains its retain, recall and reflect interface and the project's argument for memory that updates an understanding of a user rather than replaying conversation history. The speaker contrasts that approach with transcript replay, RAG retrieval and knowledge graphs.

The API examples use a bank ID and a content string to retain information, then recall stored memories. Banks organize memory by user, project or agent. Reflect connects to observations, mental models and knowledge pages, though the speaker leaves its internal mechanics to the linked documentation and paper.

The self-hosted setup overview covers a Docker container with an included database, persistent storage, an API on port 8888 and a web UI on 9999. Other paths include external Postgres, Kubernetes with Helm and embedded Python without a separate server. Local model providers include Ollama and LM Studio; the installation guide also covers Windows and air-gapped setups.

The speaker quotes the README's January 2026 LongMemEval comparison and its independent-reproduction note, but does not verify those claims. Per-model accuracy, latency and cost appear on the linked benchmark site. This is a reading of the repository and documentation: the presenter did not install, run or benchmark the software.