Hindsight
Also known as: vectorize-io/hindsight, Hindsight agent memory
Most AI agents forget everything when a session ends. Hindsight fixes that by providing a retain/recall/reflect API that agents can call to store information and pull it back on future runs. Under the hood it does not just save raw messages: it extracts atomic facts, resolves entities (so 'Alice' and 'Alice Chen' become the same person), and builds a knowledge graph that ages and consolidates over time.
The project became one of the fastest-growing AI repositories in September 2026, gaining over 4,500 GitHub stars in a single day. Its appeal is practical: it integrates with most major agent frameworks (OpenAI Agents SDK, CrewAI, LangGraph, Flowise, smolagents, Aider, Cursor, Claude Code, and others) via thin adapters, so builders can add persistent memory without rebuilding their agent architecture.
On the BEAM 10M benchmark, which tests large-scale agent memory retrieval under hard context limits (the benchmark measures how well a system answers questions that require recalling facts from very long interaction histories), Hindsight scores 64.1% compared to 40.6% for the next closest alternative. The practical use cases include coding agents that remember project conventions across sessions, support agents that recall a customer's history, and multi-agent systems where a planner and reviewer share what they learn.