Hindsight is a memory system for AI agents from Vectorize.io, designed to help agents learn from accumulated information rather than merely recall conversation history. It organizes memories into world facts, experiences, observations, and mental models, storing them in isolated banks for users, agents, or projects. Its three main operations are retain, recall, and reflect: retain extracts facts, entities, relationships, and temporal data with an LLM; recall combines semantic vector search, BM25 keyword matching, entity and temporal graphs, and time filtering, then applies reciprocal-rank fusion and cross-encoder reranking; reflect performs deeper analysis to form connections and answer questions requiring more than lookup. The system also consolidates evidence-backed observations, supports continuously updated mental models and knowledge pages, detects input language, and offers an optional memory-defense policy that can redact or block secrets and personally identifiable information using 45 patterns. The repository reports state-of-the-art performance on the LongMemEval benchmark, with independently reproduced results from research collaborators at Virginia Tech and The Washington Post, while noting that other scores are self-reported by vendors. Hindsight supports more than 25 LLM providers, including hosted, local, OpenAI-compatible, and gateway configurations, and integrates with Python, Node.js, Go, REST, CLI, MCP, coding agents, agent frameworks, and no-code tools. Deployment options include Docker, pip, Kubernetes, embedded Python, external PostgreSQL with pgvector, Oracle AI Database 23ai, and Hindsight Cloud, which offers managed infrastructure, backups, collaboration features, usage-based billing, and a stated 99.9% uptime SLA. Its documentation also describes Prometheus monitoring, webhooks, admin operations, multilingual memory handling, and an MIT license.
