Notes from the build
The longer version of what the changelog records: what we measured, what we got wrong, and what we changed. Every number links back to something you can check.
- Benchmarks7 min read
Our A/B test measured a background job, not our change
A benchmark said our new retrieval signals added 5.5 points of accuracy. The arm with fewer signals, run last, scored best: a background job was still enriching the corpus. The valid re-run, its 2.1-point noise floor, and the three checks we now run first.
- Security6 min read
Agent memory now strips credentials before it stores anything
Agents paste .env files, curl commands and connection strings into memory, and memory is built to travel. Kognite now removes credentials on write. How we measured false positives on real text, and what it still doesn't catch.
- MCP · Onboarding7 min read
Your MCP server's registry URL shouldn't need a key
Our MCP registry listing advertised a URL with an API-key placeholder, so every install needed an account before the client could connect. Hundreds of clients a day hit it and left. The fix was one line, because the gateway already spoke OAuth.
- Benchmarks8 min read
We corrected our own memory benchmark
An early chart put our LoCoMo score next to a 91.8% Mem0 figure from a different harness. We replaced it with a same-harness run where both numbers are lower and true. What the metrics mean, how to check them, and what they don't tell you.
Posts and changelog entries share one RSS feed.