Relevance AI Remembers You. It Does Not Remember Last Night's Run.

Every so often a platform ships a memory system good enough that the "does it remember" question gets interesting. Relevance AI is one of those platforms. Its documentation describes memory in unusual detail, the memories are inspectable and deletable, and the system genuinely learns. So the honest answer to "does Relevance AI remember between runs" is not the usual flat no. It is: it remembers you. What it does not remember is the run.

That distinction decides whether your scheduled agents get smarter every night or merely better-mannered.

What Relevance AI's memory actually covers

Relevance AI ships three separate memory systems, and the docs are unusually clear about each one.

Agent Memory is the long-term layer. The documentation describes it as "how the agent learns about you across conversations": it stores your preferences, corrections, project context, and references. Tell the agent your reporting format once and it keeps it. Correct a mistake — "don't include pricing in the first email" — and it saves the feedback so it does not repeat it. You can scope this memory to a whole project or a single user, review every entry in the Overview tab under Context & Memories, and delete anything that goes stale. The agent checks this memory at the start of every conversation.

Short-term memory, which Relevance calls metadata, is the working layer inside a single conversation. You define fields — statuses, extracted values, categories — and the agent fills them as it works. It exists so you can filter, search, and analyze tasks after they run. In a multi-agent Workforce, the metadata stretches across the whole run, but only while every agent is configured to "continue same task" rather than start a new one.

Chat Memory lives in the Super GTM chat. It lets you ask about your own recent conversations — "what did I compare X against last week" — and get a contextual answer. The docs state the limits plainly: memory stays within the same project, and you only ever see your own history, never a teammate's.

Notice the shape of all three: they remember the operator. Preferences, corrections, context, conversation metadata, your chat history. A memory of you.

The scheduled run is a new conversation

Relevance AI supports up to ten active schedules, each firing a saved prompt on a recurring basis. The documentation describes what happens when one fires: the agent receives the saved message "exactly as if you had typed it into a conversation." Integrations, skills, files, and memory are all applied, "just like a normal Super GTM interaction."

Read that sentence the way an automation operator should. "Exactly as if you had typed it into a conversation" means the run is a new conversation. It opens with your Agent Memory loaded — the preferences, the corrections, the project context — and with none of the previous run's content. The transcript is gone. So is the record of what the agent decided, and the trace of the approach that failed twice before the third one worked.

This is the gap, and it is specific to scheduled work. A conversational agent needs to know who you are and how you like things. A scheduled agent needs to know what happened last time it ran. The nightly lead sweep needs yesterday's dedup decisions: the records it merged, the ones it flagged as ambiguous, why it left the hard cases alone. The weekly report needs to know what last week's edition already covered. The recurring audit needs the dead ends, not just the conclusions. None of that is a preference or a correction. It is the agent's own operational history — and it is the one thing none of Relevance AI's three memories is shaped to hold.

The metadata workaround has a ceiling

The standard fix is short-term memory plumbing: define fields for run outcomes, write them at the end of each run, read them back at the start of the next. Operators do this constantly, and it works — the way a spreadsheet works as a database. But say the quiet part out loud: you are now maintaining a memory system — write discipline at the end of every run, read discipline at the start, pruning when fields go stale, contradiction handling when two runs disagree, and retrieval logic that decides which of last month's notes actually matter tonight.

And the memory is platform-locked. It lives in Relevance AI's conversation layer, which means the coding agent on your laptop cannot see it, the scheduled agent you run on another platform cannot read it, and every tool outside Relevance AI starts the relationship from zero. You fixed the amnesia inside one building and left it standing everywhere else.

Give the run its own memory

The fix is a memory layer that sits outside any single platform and remembers the work, not just the worker. Vilix AI is cloud-hosted, so there is zero infrastructure to provision or maintain — the part a scheduled operator cares about most, since the runner is already someone else's machine. It connects over MCP, so the same memory follows your agents across Relevance AI, your coding tools, and your other scheduled runners: one memory, every tool, every device.

And it keeps full conversation history, not extracted facts. That is the difference between a note that says "check the staging ledger first" and the actual transcript of the night the agent learned that rule — the failed attempts, the reasoning, the correction made mid-run. When the next run starts, the agent does not just know the rule. It knows why.

Your data stays yours: export everything or delete it at any time, in a portable format. The free plan never expires, and the 7-day Pro trial does not ask for a credit card.

Relevance AI did the honest thing. It built memory that learns the operator, documented exactly what it covers, and scoped it clearly — which is more than most automation platforms manage. For conversational work, that memory is genuinely good. For scheduled work, the run is the unit that needs remembering, and no preference store covers it. Keep Relevance AI's memory for the operator. Give the 6 a.m. run a memory of what the 6 a.m. run did yesterday, and "exactly as if you had typed it" stops being a fresh start every morning.