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