Hindsight took the top of GitHub trending with a memory system that learns, not one that recalls
Hindsight added 1,607 stars in a day to sit at number one on GitHub trending, at 27,710 total. It is an agent memory system from Vectorize.io, MIT licensed, and the pitch is a distinction most memory projects skip: recalling conversation history and learning from experience are different jobs, and almost everything shipping today does the first one.
The structure follows from that. Instead of one undifferentiated pile of embedded chunks, memories land in typed buckets — world facts, experiences, observations, mental models — and three operations move things around: retain, recall, reflect. Retrieval runs four strategies in parallel, semantic vector search plus BM25 keyword plus a graph over entity and temporal links plus straight temporal filtering, because any one of those alone fails on a predictable class of question. Storage is Postgres with pgvector, or Oracle AI Database 23ai if you're an enterprise. Memories live in isolated banks. Clients ship for Python, TypeScript, Go and the CLI, and there's an MCP server.
The operational detail worth noting: it talks to 25+ LLM providers, and four of those are subscriptions rather than API keys — claude-code, openai-codex, cursor, github-copilot. If you already pay for Claude Pro or ChatGPT Plus, the memory layer runs on that seat with no separate API bill. For anyone prototyping a memory system on their own machine, that removes the main reason not to.
On the numbers, be a little careful. Hindsight claims state of the art on LongMemEval and calls itself the most accurate agent memory system ever tested, and it does something rare and good here — its results were independently reproduced by Virginia Tech's Sanghani Center and The Washington Post, while every competing score in the same chart is vendor self-reported. But the comparison table is labeled "as of January 2026," and this is late September. Eight months is several eternities in this particular corner, and the live board at benchmarks.hindsight.vectorize.io is the number to trust over the README image.
Repo: https://github.com/vectorize-io/hindsight — docs at https://hindsight.vectorize.io/ — the paper is arXiv 2512.12818.
← Back to all articles
The structure follows from that. Instead of one undifferentiated pile of embedded chunks, memories land in typed buckets — world facts, experiences, observations, mental models — and three operations move things around: retain, recall, reflect. Retrieval runs four strategies in parallel, semantic vector search plus BM25 keyword plus a graph over entity and temporal links plus straight temporal filtering, because any one of those alone fails on a predictable class of question. Storage is Postgres with pgvector, or Oracle AI Database 23ai if you're an enterprise. Memories live in isolated banks. Clients ship for Python, TypeScript, Go and the CLI, and there's an MCP server.
The operational detail worth noting: it talks to 25+ LLM providers, and four of those are subscriptions rather than API keys — claude-code, openai-codex, cursor, github-copilot. If you already pay for Claude Pro or ChatGPT Plus, the memory layer runs on that seat with no separate API bill. For anyone prototyping a memory system on their own machine, that removes the main reason not to.
On the numbers, be a little careful. Hindsight claims state of the art on LongMemEval and calls itself the most accurate agent memory system ever tested, and it does something rare and good here — its results were independently reproduced by Virginia Tech's Sanghani Center and The Washington Post, while every competing score in the same chart is vendor self-reported. But the comparison table is labeled "as of January 2026," and this is late September. Eight months is several eternities in this particular corner, and the live board at benchmarks.hindsight.vectorize.io is the number to trust over the README image.
Repo: https://github.com/vectorize-io/hindsight — docs at https://hindsight.vectorize.io/ — the paper is arXiv 2512.12818.
Comments