Token budget for EVERY memory model call. A finished answer stops at EOS on
its own; a maxTokens cap can only ever truncate an UNFINISHED answer — there is
no case where capping output improves it (a one-word verdict already stops; a
long section that gets cut is corrupted). This is a single high backstop against
a pathological non-terminating decode, set far above any real output. Never set
a per-call cap below this.
Token budget for EVERY memory model call. A finished answer stops at EOS on its own; a maxTokens cap can only ever truncate an UNFINISHED answer — there is no case where capping output improves it (a one-word verdict already stops; a long section that gets cut is corrupted). This is a single high backstop against a pathological non-terminating decode, set far above any real output. Never set a per-call cap below this.