Technical whitepaper

Memlake: agent memory on object storage

How Hindsight Memlake, a memory engine written in Rust, keeps agent memory in object storage: searchable in milliseconds, durable before it is acknowledged, and built for billions of memories. 35 pages, with every result labelled as measured, modelled or projected.

Why agent memory needs its own store

The workload, what we measured on Postgres, and why a scale-out relational store is not the answer.

Architecture

Storage layout, the write, index and query paths, and the invariants that hold them together.

Four searches over one copy

Vector, keyword, graph and time retrieval from a single copy of each memory, with no links table.

Durability and correctness

Hardware failures, silent data errors, garbage collection, backups, restores and TLA+-verified protocols.

Evaluation

Head to head with Postgres, bank-size and fleet scaling, the price of object storage, and retrieval quality.