hindsight: Not an AI that remembers, but an AI agent that learns

While most agent memory systems focus on restoring conversation history, Hindsight focuses on building agents that learn over time. It overcomes the limitations of RAG and knowledge graphs, achieving SOTA performance on long-term memory tasks.

1 / 5

Through three core actions—retain, recall, and reflect—it stores and retrieves information to generate responses tailored to user preferences. It delivers independently verified accuracy on the LongMemEval benchmark and is in production at Fortune 500 companies.

Using a Wrapper approach that encloses existing LLM clients, you can add memory functionality with just two lines of code. It supports over 25 LLM providers, including OpenAI and Anthropic, as well as local models, and offers clients for Python, Node.js, Go, and more.