Vectorize · 2026-03-15 · notable
Hindsight — agent memory that learns
MIT-licensed agent memory system inspired by human cognition: retain/recall/reflect operations, separate networks for facts vs experiences vs opinions, and top benchmark scores on LongMemEval.
Agent memory modeled on how humans actually remember — separating facts from experiences from opinions, with top benchmark scores to prove it.
Key specs
| License | MIT |
|---|---|
| GitHub stars | 9k+ |
What is it?
Hindsight is an open-source memory system for AI agents from Vectorize, trending on GitHub since mid-March 2026. Unlike flat vector stores, it organizes memories into distinct networks for world facts, experiences, observations, and opinions, separating evidence from inference. Agents interact through three operations: retain() to store, recall() to search, and reflect() to reason across memories and form new insights.
How does it work?
Recall uses four parallel retrieval strategies: semantic search, keyword matching, graph-based traversal, and temporal ordering. Memory is time-aware and contextual, so the system can distinguish between current and outdated information. It integrates via Docker containers, Python packages (with an embedded server option), or Node.js/TypeScript SDKs, and adds to existing agents with as few as 2 lines of code. It achieves top scores on LongMemEval, the standard benchmark for conversational AI memory.
Why does it matter?
Most agent memory solutions dump everything into a vector store and hope retrieval works. Hindsight's structured approach — treating memory like a cognitive system with different stores for different types of knowledge — produces measurably better recall. It is already deployed at Fortune 500 companies and AI startups.
Who is it for?
Agent framework authors, developers building long-running AI assistants.
Try it
pip install hindsight-server