Hindsight

Hindsight is an open-source agent memory system from Vectorize that helps AI agents learn over time instead of merely remembering chats, with more than 52,000 stars on GitHub. It reports state-of-the-art results on the LongMemEval long-term memory benchmark.

Overview

Most agent memory add-ons simply stuff conversation history into a vector store and retrieve it with RAG. Hindsight is built around a different model: agents store experiences with a Retain operation, search them with Recall, and use Reflect to generate disposition-aware answers that improve as more evidence accumulates. Memories live in organized memory banks and are distilled into observations and mental models, so the agent builds durable knowledge about users, projects and preferences rather than replaying raw transcripts.

You can self-host Hindsight with Docker, pip, a Helm chart for Kubernetes, or even embedded in Python with no server at all. It works with more than 25 LLM providers, from OpenAI, Anthropic and Gemini to fully local options like Ollama and LM Studio, and offers official clients for Python, Node.js, Go, a CLI and a REST API.

Key Features

• Retain, Recall, Reflect – three core operations designed for learning agents, not just chat replay.
• Memory banks and mental models – structured, per-user or per-project knowledge that consolidates over time.
• Top benchmark accuracy – state-of-the-art reported performance on LongMemEval, independently reproduced by research collaborators.
• Two-line LLM wrapper – wrap an OpenAI or Anthropic client via LiteLLM and memory is stored and recalled automatically on every call.
• 60+ integrations – Claude Code, Cursor, GitHub Copilot, LangGraph, CrewAI, LlamaIndex, the OpenAI Agents SDK and many more, plus an MCP server.
• Flexible deployment – Docker, bare-metal pip, Kubernetes Helm charts, or an embedded Python mode with no server required.
• Any LLM provider – 25+ hosted and local providers, including subscription-based access through Claude Code, Codex and Copilot.

💳 Pricing

Hindsight is free and open source under the MIT license for self-hosting. A managed Hindsight Cloud option is available with usage-based billing and free credits to start, for teams that prefer not to run the server themselves.

Who Is It For

Developers building AI agents, copilots and assistants that need reliable long-term memory across sessions, and teams that want that memory layer self-hosted, inspectable and provider-agnostic.