That sentence is the whole strategy behind Wallaby, and this post is us introducing ourselves properly.
Pick is judgement. Today Wallaby is a curated inference layer for open-weight models: one API, prepaid credit, per-token metering, itemized statements. The shelf is small on purpose (one model today, Kimi K3) because a model earns its place by being something we run in production ourselves. Fewer models, better choices.
Infrastructure is the longer bet. Models are becoming commodities; the durable value sits around them — agents, context, memory, reliability. We build that layer in a specific order: hit a real problem running our own agents, build the fix, run it in production, publish what happened, then productize. Build, use, measure, publish.
The first instance of that loop is memory.
Agents forget. And when they do remember, you usually can't see what they remember, where it came from, or why it is still there. Our own agents hit this wall daily, so we built the discipline we needed: memory as plain files, with rules. Every memory carries a receipt — a date and a source. Memory files have size thresholds and age out when idle. A secrets scan runs before anything lands. It is not a product launch; it is the rulebook we run on, and it is public: wallaby-agent-rules.
The research log behind it is public too, on our blog — what we audited, what broke, what we changed, with the numbers left in.
What happens next: we keep enriching this memory system in public. More rules, more measurements, more honest failure logs — the repo and the blog both keep growing over the coming weeks.
If you think agent memory should be infrastructure you can inspect rather than a black box you trust, come find us. Open an issue on the repo, or leave a comment below — we read everything.
One promise for this account: no posts about what should work. Only what actually happened.
Top comments (0)