Unconfirmed: a self-hosted memory system for AI agents called Hindsight is being passed around developer timelines, with claims about both its popularity and its accuracy. Automatica has not verified the star counts or the benchmark result, and both trace back to the project and its promoters.
The account RoundtableSpace wrote that the project "picked up 1,668 GitHub stars in a day on its way to 22.1K", organises memory into "world facts, experiences, observations and mental models, modeled on human memory", and "claims state-of-the-art accuracy among agent memory systems on the LongMemEval benchmark". The same post says it plugs into "60+ tools", runs locally under Docker or pip, and is MIT licensed.
Note the phrasing of the accuracy line: it is a claim the project makes about itself. LongMemEval is a public benchmark, so the result is reproducible in principle by anyone who wants to run it, which puts this in a different category from a vendor's private numbers.
A second account, cai_smart, described the design in more detail and drew the distinction the project is built on: "Most agent memory systems are a conversation log with a search box bolted on. This one is built around the claim that recall isn't the same as learning." It sets out three operations — retain, recall and reflect — with clients for Python, Node, Go and a command line.
Star counts are the weakest evidence in any of this. They are easily promoted, they say nothing about whether the software works, and the figure here has not been checked against the repository.
What would settle the substance: an independent LongMemEval run by someone other than the authors, and reports from developers who have kept it in production for more than a few days. Both are easy to produce, which is why the next week of posts about this project will be more informative than this one.