What Waku got right

Last month a product called Waku Memory appeared with a landing page that could have been ours: your portable memory, one memory that follows you across AIs. Our first reaction was the normal one. Our second reaction was more useful: we read everything they shipped, wrote down what they did better than us, and spent a week closing the list.

This post is the list. Competitors usually get subtweets; the ones who teach you something deserve citations.

1. Deterministic capture, with consent done right

Smara's original answer to “how do memories get in” was a Skill: teach the assistant a memory discipline and trust it to call the tools. It works — but it depends on the model remembering to remember. Our own beta data said the quiet part: most people who connected an assistant and stopped there never saw their memory do anything.

Waku's answer is hooks. Their setup installs a small hook in Claude Code that captures each turn deterministically — no model cooperation required. And before turning it on, it tells you exactly what will be sent and what never will, then asks. That consent screen is the best privacy UX we have seen in this category.

So we built smara-capture: a consented, async hook that records each Claude Code turn — your message, the reply, tool names— and never file contents, commands, tool inputs, or reasoning. The never-list isn't a promise in a README; it's a test in our suite that plants a secret in a tool input and fails if it leaks. The consent flow shows everything first and requires an explicit yes, because Waku set that bar and it is the right bar.

The difference is what happens after capture: Smara's background consolidation distills the turns into dated facts, a knowledge graph, and an auditable history — the hook is a front door to the same memory your other assistants share.

2. Onboarding is one command, not three steps

npx waku-memory setupis a better first minute than our paste-this-config-then-download-this-file page, and there's no pretending otherwise. So: npx smara setup now exists — it registers the connector in Claude Code, installs the memory Skill, offers capture (through its own consent flow), and configures Cursor, asking before every write. The chat-app flow keeps its three friendly steps; developers get the one-liner they deserve.

3. “Files you own” is a feeling, not just a feature

Smara has had export-everything from early on — facts, graph, events, the full audit trail, one API call. Complete, and abstract. Waku says “files you own, on a machine you own” and you feel it in your chest.

They're right about the feeling, so we built the feeling on top of the substance: npx smara pullmaterializes your entire memory as readable files on your disk — one markdown per fact with its confidence, validity window, and provenance in the frontmatter; your episodic ledger month by month; your graph as JSON; the complete audit trail including every edit your dreams made. Rebuilt any time you run it. It's a projection of the same memory, which is the point: the files are yours because the memory was always yours.

What we didn't adopt

Credit cuts both ways, so for the record: we didn't adopt wall-clock dating (a memory system that resolves “last week” against today instead of when it was said will misremember your life politely), write-time-only curation (contradictions and duplicates deserve a nightly process with receipts, not a single gate), or silence about accuracy (we publish benchmarks, including the runs that went down).

Good competitors make the category legible and keep everyone honest. Waku made us better in a week. We hope this page returns the favor.