Hindsight ships v1.0 with Memlake, a Rust memory engine on object storage

An AI agent that cannot remember what the user told it last Tuesday has to be reminded of everything on every call, and the memory layer that fixes it has quietly become the shelf most teams buy from rather than build. Vectorize's Hindsight is one of the four open-source libraries that has been on that shelf for a year, and it has just crossed its v1.0 milestone.

The release went out on the project's GitHub as v1.0.0, after a run of v0.10.x point releases. The repository, vectorize-io/hindsight, carries 49,424 stars under the MIT licence. The company lit up a product page for Hindsight Memlake with the release, a Rust storage engine built for the shape of data an agent writes, as the paid half of a Hindsight Enterprise plan.

What Vectorize's own benchmarks claim

The vectorize.io landing page shows a bar chart titled "The learning difference, measured" with four rows. Hindsight scores, each from Vectorize's own run, are 94.6% on LongMemEval-S (next-best system 74.0%), 92.0% on LoComo (next-best 80.3%), 86.6% on PersonaMem (next-best 84.4%), and 85.7% on PrecisionMemBench with "no published comparison" labelled on the chart.

The underlying leaderboard, agentmemorybenchmark.ai, is itself run by Vectorize, so the comparison is a vendor's benchmark of a vendor's library against named competitor systems. Treat the gap figures as what Vectorize's own run shows, not as a cross-vendor finding.

What Memlake adds for enterprise

The Memlake product page describes a storage engine with four parts: each organisation gets its own object-storage bucket, writes land as immutable log objects, indexers fold the log into segments that are written once and never modified, and the manifest and head pointer coordinate through conditional writes, so there is no database, lock service or leader beside the bucket itself.

The page states the target shape is "100 million memories at under 100 ms p99, every write durable before it is acknowledged, and four kinds of search over one copy, in a bucket you control." Readers see whole versions because nothing is overwritten in place, and a bank's index lives on one server at a time, which another picks up when a server disappears.

The open-source Hindsight stores memory in Postgres by default and the project page states that is the right choice for most deployments. Memlake exists for the case where it stops keeping up.

What a v1.0 brings to the stack

The release notes list the usual mix of fixes and small features, including per-integration User-Agent headers, Windows hooks fixes for Claude Code and Cursor, native memory provider registration for OpenClaw, and a shared-bank page scope for coding agents. The one breaking change is the removal of the agent-sdk package and the agent_knowledge_* tools, which the release calls refactor(openclaw)!.

The practical read for a team adopting Hindsight today is that v1.0 is a stability stamp on a library that already had integrations for Claude Code, Codex, Cursor, the Pydantic AI framework, LlamaIndex and Hermes. The vendor-published benchmark edge is loud enough to put Hindsight on the short list, with the caveat that the gap numbers come from Vectorize's own run. A team picking a memory library this quarter should run its own three-question check: does the integration for its coding agent handle sessions the way it expects, does the on-disk layout fit the compliance picture, and do the benchmark scores hold on the team's own memory shape when the vendor's rank is not the one being measured.

Source

Vectorize website, the Hindsight v1.0.0 release, the Hindsight Memlake page, and the benchmark subsite.