Hindsight

Overview

Hindsight is an agent memory system designed for AI agents that need long-term memory across sessions. It extracts conversations into typed facts, consolidates them into observations, and supports recall and reflective reasoning over memory banks.

Key Features

• Typed memory extraction turns conversations into world facts, experience facts, observations, mental models, and knowledge pages rather than storing raw transcripts.
• Multi-strategy recall runs semantic, keyword BM25, graph, and temporal searches in parallel, fuses results by rank, reranks with a cross-encoder, and cuts to a token budget.
• Observation consolidation deduplicates overlapping facts, grounds observations in evidence, keeps a history of refinements, and treats affected observations as stale when new memories are not yet consolidated.
• Reflect runs an agentic loop over the bank, gathers its own evidence, checks mental models and knowledge pages before raw facts, and cites retrieved sources.
• Memory banks are isolated per user or agent and can be configured with mission, directives, and disposition to shape reflect behavior.
• Integrations include Python, TypeScript, Go, CLI, MCP, personal agents, and 20+ coding-agent harnesses.

Use Cases

• Coding agents can use one install to wire 20+ harnesses to a per-repo memory bank.
• Personal agents such as Hermes and OpenClaw can use Hindsight as their memory provider.
• AI agents can use Hindsight to recall context across sessions, reason over time expressions, and consolidate beliefs from many facts.
• Teams can run Hindsight Cloud or self-host it for production memory needs.

Getting Started

• Install and get running in 60 seconds via the Quick Start.
• Start a server using Docker Compose, Helm, or pip, then connect a client.
• Use Python, TypeScript, Go, CLI, MCP, or plain HTTP clients.
• Use the Python embedded option when no server is required.

Deployment & Requirements

• Hindsight Cloud is a fully managed option.
• Embedded deployment has zero setup and runs locally.
• Self-hosted deployment supports Docker Compose, Helm/Kubernetes, and bare-metal pip.
• Memory banks use Postgres with pgvector, and users can bring their own cluster connection string.

Before You Adopt

• License: MIT. Review its terms before using, modifying, or distributing the project.