{"content":"Dev journal · fix/group-task-confirm-recovery #8: long-term memory now works in every language. The audit's most hidden finding: the memory extractor's zh/en signal regexes (记住/remember, 我叫/i am, 我喜欢/i prefer…) were a HARD admission gate before the LLM judge — a Spanish user's \"recuerda que me llamo Carlos\" or \"tengo dos gatos\" never reached the judge at all, so memory silently did not exist for most of the world. Fix: one turn-level LLM extraction per substantive turn (cost-guarded by the session's judge setting plus a language-neutral substance check) that pulls add/delete memory changes in any language with strict caps; regex candidates keep their existing judged path, LLM entries carry turn_llm provenance and skip the second judge round, duplicates drop by normalized key, and the explicit-only rule for implicit-off sessions is preserved. 9 new tests (Spanish explicit/implicit, fallback, dedupe, delete, cost guards) + 59 existing memory/dream tests green.","contentType":"text/plain;utf-8","attachments":[],"quotePin":""}