WHAT IS IT?
Hindsight is a memory system for AI agents built by Vectorize. Where most solutions merely store and retrieve conversation history, Hindsight learns: it extracts facts, entities and relationships, organizes them into memory networks inspired by human cognition, and generates new conclusions through reflection. The whole thing is self-hostable and posts state-of-the-art results on the LongMemEval benchmark.
WHY IS IT INTERESTING?
- Three operations, one simple mental model: retain stores information with automatic extraction of facts, entities, relationships and temporal data; recall retrieves the relevant memories; reflect analyzes existing memory to form new connections.
- Multi-strategy recall: four approaches run in parallel: vector similarity, BM25 keywords, entity graph traversal and temporal filtering. No need to choose between RAG and a knowledge graph, Hindsight combines both.
- Cognition-shaped memory: memories are organized into world facts, experiences, opinions and observations, rather than piled into a flat index of vectorized chunks.
- Two-line integration: an LLM wrapper adds memory to an existing agent without a rewrite. OpenAI, Anthropic, Gemini, Groq and Ollama are supported on the provider side.
- Flexible deployment: Docker container backed by PostgreSQL, or embedded in-process Python mode with no separate server. Python and TypeScript clients, per-user memory isolation, MIT license.
USE CASES
- Give a chatbot long-term memory, with per-user isolation for personalization
- Make an agent improve from one run to the next (support, code review, ops) instead of starting from scratch
- Replace a homegrown RAG pipeline when plain vector search hits its limits
- Build persistent user profiles fed by conversations
- Experiment with agent memory research: an arXiv paper accompanies the project