Hindsight

vectorize-io/hindsightReviewed by AI

What it does

A memory system for AI agents that you run yourself. Your program sends text in (retain), asks questions later (recall), or asks for a worked-out answer built from everything stored (reflect). Facts, experiences and background summaries live in a PostgreSQL database, and there are clients for Python, Node, Go and a CLI, plus ready-made hooks for coding agents, agent frameworks and MCP. For developers building agents or chatbots that should remember. Start with one Docker command or `pip install hindsight-api`, plus a key for one of 25+ model providers; local models also work.

How easy it is to install

C · install 1 or 2 tools · Windows

Windows C
macOS C
Linux C

C on Windows, Mac and Linux: the shortest path is `pip install hindsight-api`, one environment variable with a model key and one command, with the database bundled (the README's platform table marks the embedded database as supported on all three). The Docker route is one long command instead. The release files are only a small command-line client for Linux and Mac, and Intel Macs need the separate `hindsight-all-slim` package.

Numbers on GitHub

Stars: 56,983on Oct 31, 2026
Downloads: 56,4601,939 this week
Growth: +2,878stars in 5 days

Stars over time

Facts

Runs on: Windows, macOS, Linux
You need to install: Python
To build from source: Python, Node.js
Disk space: 200 MB–800 MB · estimate
Accounts or keys: OpenAI, or any of the 25+ supported model providers (a local Ollama or LM Studio needs no key)
Cost: Free, API costs extra · The code is MIT and self-hosting is free, but every retain, recall and reflect calls a language model, so you pay that provider. The README says local providers (`ollama`, `lmstudio`, `llamacpp`) and existing subscriptions (ChatGPT, Claude, Cursor, GitHub Copilot) work too. The company behind it also sells Hindsight Cloud, a usage-based hosted version.
License: MIT (Permissive)
Can replace: Mem0, Zep, Supermemory
Last update: Oct 30, 2026

Risk signals (6)

Attentioneverything you store passes through a model provider
README: "Behind the scenes, retain uses an LLM to extract key facts, temporal data, entities, and relationships", and the quick start sets `HINDSIGHT_API_LLM_API_KEY` to an OpenAI key. Unless you point it at a local model, the text of your memories is sent to that provider.
AI review
Infocoding agent setup ingests your repository history by itself
README, "Coding Agents": `npx @vectorize-io/hindsight-coding-agents install all` wires "every detected agent", builds "a per-repo bank built automatically from git history and past sessions", and "Ingestion is automatic — there is no setup command."
AI review
Infosecret and personal data scanning is off by default
README, Memory Banks: Memory Defense is "An opt-in, per-bank policy that scans every retain for secrets and PII against 45 patterns". Until you turn it on, tokens and personal data in what you store are kept as they come.
AI review
Infocommand line client installed by piping a script into bash
README, "Connect a client": `curl -fsSL https://hindsight.vectorize.io/get-cli | bash` runs whatever that address returns, with no checksum shown.
AI review
Infobenchmark comparison comes from the vendor
README: "Hindsight is the most accurate agent memory system ever tested according to benchmark performance", with a chart from the project itself; the same section admits "Other scores are self-reported by software vendors." It also states, without evidence in the repository, that it "is being used in production at Fortune 500 enterprises".
AI review
Infocounting pixel in the readme
The last line of README.md is a 1x1 image: `<img src="https://umami-pixel.chris-latimer.workers.dev/?id=a8b043e6-6964-454d-80df-69b69d3f0d50&host=github.com&url=/vectorize-io/hindsight" width="1" height="1">`, which reports views of the page to the author's own server.
AI review

Evidence is shown in its original language.

What Radar could not check

What it really costs in model tokens to keep a busy agent's memory up to date.
Whether the benchmark numbers in the README reproduce on your own data.
How much disk the database uses as memories pile up.
Whether the pip and embedded-database routes behave on Windows as the platform table claims.
Whether bank isolation ("no cross-bank leakage") holds in practice when several users share one server.
What Hindsight Cloud costs once the free credits are used.