# Hindsight

An agent memory system built to let agents learn from experience over time, rather than only recall past conversations, reporting strong results on a long-term conversational memory benchmark.

Hindsight organizes agent memory around retaining, recalling and reflecting on observations, mental models and knowledge pages grouped into banks, aiming to go beyond simple retrieval-augmented generation or knowledge-graph memory by having the agent learn from what it stores. On the LongMemEval benchmark, used to assess memory system performance across conversational AI scenarios, the project reports state-of-the-art results as of January 2026, with the benchmark data independently reproduced by outside research groups.

It runs as a server, in Docker or on Kubernetes via Helm, with an API and a UI, and can also install as a package on bare metal. It connects to more than 25 LLM providers, including hosted APIs, fully local runtimes such as Ollama, llama.cpp and LM Studio, OpenAI-compatible endpoints, gateways, and existing subscriptions such as ChatGPT Plus/Pro, Claude Pro/Max, Cursor and GitHub Copilot without needing a separate API key. It supports PostgreSQL and Oracle AI Database for storage.

It is written in Python and licensed under the MIT license.

Website: https://hindsight.vectorize.io/