Introducing Hindsight Memlake: agent memory on object storage. 100M memories. Recall under 100 ms p99. Same Hindsight. Up to 47x faster. Up to 26x cheaper.

Game changer

I ran some tests on hindsight a few weeks ago and am debating whether to incorporate into my stack. Sounds like this is a total overhaul?

Agent memory only grows. An autonomous agent can write ~5M memories a year, and every recall runs vector, keyword, graph and time search over all of it. Same cluster, same data, a ~200k-memory bank: Postgres: 5.3 s p99 Memlake: 113 ms

How it works: • Object storage is the only source of truth • Only two objects ever change, via conditional PUTs (If-Match) • Servers are disposable caches: RAM + NVMe • One copy serves vector, BM25, graph and time search Written in Rust.

Also today: Hindsight is 1.0. A year in production, #1 on every dataset of the Agent Memory Benchmark, and an API that's now a promise. Still MIT, still runs on Postgres. The engineering deep dive:

Memlake runs in Hindsight Cloud, and in Hindsight Enterprise: your own AWS, Google Cloud or Azure account, or on premises. The 35-page whitepaper, with every result labelled measured, modelled or projected:

insane work we did, this is the future of memory of ai agents, petabytes of data

Nice job guys. Grats

We have been using graphiti? Is it worth to migrate ?