The First Agent Memory System
That Actually Learns
Hindsight is the first agent memory system with a mechanism for agent learning. Agents can reflect on their experiences, form opinions, and avoid repeating mistakes.
Agent Memory That Works Like Human Memory
Hindsight uses multiple memory networks to create the first agent memory system with a mechanism for agent learning. When agents reflect on their experiences, they form a better understanding of what works and what doesn't.
- World Facts
Objective facts about the agent's world. "The customer made two purchases and asked to return one item."
- Experiences
Things the agent has experienced - conversations, interactions, tool calls. "In the past, my attempts to issue a refund failed because I didn't have a refund request number."
- Opinions
Fluid perceptions of facts and experiences that can change over time with confidence scores. "I should make sure I ask the customer if they have a refund request number before I create one to avoid duplicates."
- Observations
Broader connections between facts and experiences. "Customers who ask for a refund often want a replacement and I can improve their experience by offering to send a replacement proactively."
State of the Art Benchmark Performance
Hindsight achieves 91.4% on LongMemEval, the first agent memory system to cross 90% and the highest score ever reported as of December 2025. Backed by peer reviewed research with collaborators from top academic institutions and industry experts.
Open Source
MIT Licensed
Hindsight is completely open source with an MIT license. Use it in any project, contribute to development, or fork it for your own needs. No vendor lock-in, no restrictions.
Free to Use
No licensing fees, no restrictions on commercial use
Community Driven
Built by the community, for the community
Production Ready
Battle-tested and ready for production workloads
Build agents that actually remember.
Join the open source community building the future of agent memory. Hindsight is available now on GitHub with comprehensive documentation and examples.