Hindsight

Hindsight is an MIT-licensed agent memory system built by Vectorize, Inc. that enables AI agents to retain, recall, and reflect on information across sessions. It is available as a self-hosted open-source project (deployable via Docker, Kubernetes, or pip) and as a managed cloud service. The project claims state-of-the-art performance on the LongMemEval benchmark, with a reported score of 94.6%, independently reproduced by research collaborators at Virginia Tech's Sanghani Center for Artificial Intelligence and Data Analytics and The Washington Post.

What It Is

Hindsight is a purpose-built memory layer for AI agents, distinct from traditional RAG (Retrieval-Augmented Generation) systems. Where RAG retrieves document chunks, Hindsight stores structured facts, builds consolidated observations over time, and uses a graph-based event-centric architecture to track how knowledge evolves. The core API exposes three operations: retain (store information), recall (search and retrieve memories), and reflect (generate disposition-aware answers grounded in memory). Memory is organized into banks — isolated stores that can be configured with a mission, directives, and disposition traits to shape how the agent reasons.

Architecture: Four Memory Networks

Hindsight organizes knowledge into a hierarchy of memory types:

• Mental Models — user-curated summaries for common queries
• Observations — automatically consolidated, deduplicated beliefs built from raw facts over time, with evidence tracking and freshness awareness
• World Facts — objective facts received (e.g., "Alice works at Google")
• Experience Facts — the agent's own actions and interactions (e.g., "I recommended Python to Bob")

Retrieval uses a four-strategy parallel system called TEMPR: Semantic (vector similarity), Keyword (BM25 exact matching), Graph (entity and causal links), and Temporal (time-range filtering). Results are merged via reciprocal rank fusion and a cross-encoder reranker. The graph is event-centric — entities attach to memory units rather than connecting directly to each other — which preserves historical state without requiring manual rewrites when facts change.

Deployment Model

Hindsight supports multiple deployment paths:

• Docker — single command, embedded PostgreSQL, runs on Linux, macOS (Apple Silicon and Intel), and Windows
• Kubernetes / Helm — production-grade with StatefulSet worker identity management
• Bare metal (pip) — pip install hindsight-all for Python-embedded use with no separate server
• Hindsight Cloud — fully managed, usage-based billing on token consumption with no fixed monthly fee

Client SDKs are available for Python, TypeScript/Node.js, Go, and CLI. The system integrates with over 40 agent frameworks and tools including LangGraph, LlamaIndex, CrewAI, AutoGen, OpenAI Agents SDK, Claude Agent SDK, Cursor, Cline, n8n, Zapier, and many others.

Update: v0.8.3

The latest release is v0.8.3, published on June 18, 2026. The GitHub repository was created in October 2025 and has accumulated over 16,700 stars and 975 forks as of late June 2026, with active pushes as recently as June 20, 2026. The project also claims the #1 position on the BEAM benchmark as of April 2026, per a blog post on the Hindsight site. The current documentation version is 0.8, with archived versions for 0.7 and 0.6 also available.

Security and Isolation

Hindsight includes a Memory Defense layer for security. User data isolation is handled through memory banks — each bank is fully isolated with no data leakage. Developers can create per-user banks for maximum privacy or use a single bank with tag-based filtering for cross-user analytics. Entity labels with tag: true enable controlled-vocabulary classification at retain time and deterministic SQL-level filtering at recall time, which is more reliable than scoring-based approaches for strict memory-type separation.