vectorize-io/hindsight

Hindsight is a memory system for AI agents, the software programs that carry out tasks or hold conversations on a user's behalf. The README's main argument is that most existing memory tools just recall past conversation, while Hindsight is meant to help an agent actually learn from experience over time, not only remember what was said. To explain the idea in plain terms: an AI model on its own forgets everything once a session ends. Hindsight gives it a place to store facts and to look them back up later. The README describes three core operations with simple names: retain, which saves a piece of information, recall, which searches for relevant stored memories, and reflect, which produces a considered response based on what has been stored. Memories are grouped into named banks, and extra metadata can be attached so that, for example, each user's memories stay separate. The project makes strong performance claims. It says it scored highest on a public benchmark called LongMemEval that measures memory accuracy, and notes that those results were independently reproduced by researchers at Virginia Tech and by The Washington Post, while competitors' scores are self-reported. It also states it is used in production at large enterprises. These are the authors' claims as written in the README. There are several ways to add it to an agent. The simplest is a wrapper that replaces your existing connection to a language model with two lines of code, after which memory happens automatically. For more control there is an API with Python and JavaScript client libraries. You can run the server with Docker, connect it to your own PostgreSQL or Oracle database, or run a fully embedded Python version with no separate server. It works with several model providers, including OpenAI, Anthropic, Gemini, Groq, and local options. The code is released under the MIT license, and the README links to documentation, a research paper, and a hosted cloud version.