What is Hindsight?

Hindsight is an open-source (MIT) agent memory system designed to make AI agents smarter by enabling genuine learning across sessions. Unlike traditional RAG or simple conversation history, Hindsight treats memory as a first-class substrate for reasoning.

It solves the problem of stateless agents that forget everything between interactions by providing structured, biomimetic memory with three core operations: Retain, Recall, and Reflect.

Core Features

Biomimetic Memory Networks: Organizes knowledge into World Facts, Experiences, and Mental Models (including automatic Observation consolidation).
Retain: LLM-powered extraction of entities, relationships, facts, and temporal data into canonical memory banks.
Recall (TEMPR): Hybrid multi-strategy retrieval combining Semantic (vector), Keyword (BM25), Graph (entity/temporal/causal), and Temporal filtering, fused with reciprocal rank fusion and reranking.
Reflect: Agentic synthesis using memory to generate insights, update beliefs, and support complex reasoning. Configurable via Mission, Directives, and Disposition.
Memory Banks: Isolated per-user or per-context storage with metadata support.
Multi-LLM Support: Works with OpenAI, Anthropic, Gemini, Groq, Ollama, and more via LiteLLM.

Performance

Hindsight achieves state-of-the-art results on long-term memory benchmarks, including top accuracy on LongMemEval (as of early 2026). Performance has been independently reproduced by Virginia Tech’s Sanghani Center and The Washington Post.

Installation & Quick Start

Docker (Recommended)

export OPENAI_API_KEY=sk-xxx
docker run --rm -it --pull always -p 8888:8888 -p 9999:9999 \
  -e HINDSIGHT_API_LLM_API_KEY=$OPENAI_API_KEY \
  -v $HOME/.hindsight-docker:/home/hindsight/.pg0 \
  ghcr.io/vectorize-io/hindsight:latest

API runs on http://localhost:8888, web UI on http://localhost:9999.

Python SDK

pip install hindsight-client -U

Basic usage:

from hindsight_client import Hindsight

client = Hindsight(base_url="http://localhost:8888")
client.retain(bank_id="demo", content="...")
results = client.recall(bank_id="demo", query="...")
insights = client.reflect(bank_id="demo", query="...")

Node.js/TypeScript and embedded modes are also supported.

Use Cases

Building persistent coding agents (Claude Code, Cursor integrations)
Personalized conversational agents with long-term user preference retention
Autonomous agents that learn from experience and adapt strategies
Enterprise AI workflows requiring cross-session memory and reasoning

Resources

GitHub: https://github.com/vectorize-io/hindsight
Official Docs: https://hindsight.vectorize.io
arXiv Paper: https://arxiv.org/abs/2512.12818
Hindsight Cloud: https://ui.hindsight.vectorize.io

Hindsight is actively maintained (latest commit April 2026) and used in production by Fortune 500 companies and AI startups.