token-level context compressionContextCrumb keeps the signal and shakes off the padding.
A small model scores every token before your agent spends context on it.
Agents spend context on notes, logs, tickets, docs, and tool descriptions. Those files contain useful facts, but they also carry filler phrases and repeated wording. ContextCrumb scores each token, keeps the original order, and removes low-value padding.
Agentsspendcontextnotes,logs,tickets,docs,tooldescriptions.filescontainusefulfacts,carryphrasesrepeatedContextCrumbscorestoken,keepsoriginalorder,removeslow-valuepadding.
37 input tokens25 keptsame order