Memory

Agent Ops projects learn. During every run, important facts are extracted and saved to the project's memory automatically — no "remember this" command needed — and recalled in later runs when they're relevant.

What gets remembered

• Decisions and preferences — "the CFO wants amounts in EUR", "region names follow the sales-territory list, not the country list"
• Formats that worked — the structure of a report the owner accepted without changes
• Environment learnings — which package was installed and why, so the next run doesn't rediscover it
• Data quirks — "the export always has a duplicated header row", "customer IDs changed format in March"

Why it matters for recurring work

Memory is what turns a recurring task from "same prompt, fresh amnesia" into an improving operation:

• Week 1: the weekly report task asks two clarifying questions about format.
• Week 2: it applies the answers without asking.
• Week 8: the source export changes a column name — the agent notices because it remembers the old shape, adapts, and notes the change.

Combined with the persistent sandbox (which keeps the files) memory keeps the understanding — together they give a project genuine operational continuity.

Where to see it

Each task screen has a Memory panel showing the facts saved and recalled for that work, so what the agent "knows" is never a black box. Memory is scoped to the project's workspace — projects don't leak knowledge into each other.