Memory bloat, compaction loss, and a retrieval-first path: ~32% less token spend on the AppWorld dev split — without dumbing the agent down.

Developers who actually ship with LLMs know one truth by heart: the context window is not free. Every extra thousand tokens nudges the invoice up and the latency out.

If you run OpenClaw (an agent stack that leans hard on long-horizon sessions), that anxiety gets concrete fast. Picture this: last week you spent two hours with your agent debugging production — logs, configs, experiments — and burned through 30k tokens of back-and-forth. This week you pick up where you left off, and the agent answers: Hi! Which refactor are we talking about?

So you spend a few thousand tokens re-explaining context. The model spends a few thousand more re-understanding. And you still might not land the same mental model you had last Tuesday.

Those 30k tokens? Mostly gone.

That is not a one-off glitch.…