@jlongster A pattern you can implement is a turn router. Use a cheap model like gpt oss 20b to classify the turn then retrieve relevant memory artifacts and pass it to the main model. That way the memory is always contextual. Youโd make two model calls per turn but gpt oss is dirt cheap.
@ryancarson@openclaw I donโt think it requires opus 4.6 grade intelligence all the time. If you built in a way to route action tasks to kimi and planning to opus; you could bring that down to $10-25/day
Not so hot take.
Humane gave up too soon;
Imagine if you could hook up your claw to your humane device, chat with it, get updates etc etc.
Not sure if they gave up or their VCs wanted out.
"Tonight at 6:24 P.M. EST, for the first time in over 50 YEARS, America is going back to the Moon! Artemis II, among the most powerful rockets ever built, is launching our Brave Astronauts farther into Deep Space than any human has EVER gone. We are WINNING, in Space, on Earth, and everywhere in between โ Economically, Militarily, and now, BEYOND THE STARS." - President Donald J. Trump ๐บ๐ธ
Iโve been on the same codex session for 4 days - multiple context compressions
And itโs still going; I think itโs better than starting with a fresh session.
NGL openai cooked with codex + gpt 5.4 xhigh.
The cursor disclosure today is really pushing the envelope on what โproprietary modelโ means.
Can we now just RL kimi; slap on a new name and raise a billy? ๐ค