Hindsight overview: self-hosted memory for AI agents

Hindsight is an MIT-licensed, open source memory system for an AI agent. This overview explains its memory architecture and deployment options, including a self-hosted server backed by PostgreSQL. The speaker describes support for Linux, macOS and Windows, including Apple Silicon and Intel Macs, with a lighter build recommended for older Intel hardware.

The explanation centers on retain, recall and reflect. Retain uses a language model to extract facts, entities, relationships and dates. Recall combines semantic search, BM25 keyword search, graph retrieval and temporal filtering before reranking results within a token budget. Reflect reasons across stored memories. Observations consolidate supporting evidence and change as new information arrives; mental models store answers to recurring questions that an agent can read without another model call.

Deployment coverage includes Docker with embedded or external PostgreSQL, pip installation, an embedded Python option and a Kubernetes Helm chart. The speaker also describes Ollama and LM Studio support for a local LLM, plus an LLM wrapper and MCP endpoints for connecting agents. A coding assistant integration builds repository memory from Git history and past sessions.

The reported LongMemEval score of 91–94% is a claim presented in the video. The speaker also covers opt-in Memory Defense and cautions that Hindsight's complexity may exceed the needs of simple automation workflows.