Hindsight: Agent Memory That Learns
Source: https://github.com/vectorize-io/hindsight Author: Vectorize.io Date: 2025-2026
Summary
Production agent memory system using three memory pathways (world memories, experience memories, mental models) and three operations (retain, recall, reflect). State-of-the-art on LongMemEval. Deployed at Fortune 500 enterprises. Different from the arxiv paper of the same name.
Key Claims
- Three memory types: world facts, agent experiences, mental models (inferred from reflection)
- Three operations: retain (extract entities/relationships/temporal), recall (parallel semantic+keyword+graph+temporal), reflect (analyze → new insights)
- Outperforms RAG and knowledge graphs for agent memory
- State-of-the-art on LongMemEval benchmark (verified by Virginia Tech and Washington Post)
- Production deployments at Fortune 500 enterprises and AI startups
- Supports Python, Node.js, REST, CLI; Docker deployment
Concepts
- Agent Memory — production implementation of multi-type memory system