Tray.ai Remembers the Conversation. Your Scheduled Agent Still Starts From Zero.

Tray.ai ships with memory. That sentence is true, and it is also the source of most of the confusion about what the platform does between scheduled runs. In June 2025, Tray.ai released a major update to its Merlin Agent Builder that added what the company calls "maximum short- and long-term memory." Agents built on the platform can track session history and refer back to prior conversations automatically. Sliding context windows keep even long, complicated threads coherent. If your question is "will my agent remember what the user told it last week," the answer is yes.

If your question is "will my scheduled agent remember what it did last night," the answer is not the one you want. And the difference between those two questions is the difference between an agent that is pleasant to talk to and an agent that gets smarter at its job.

Tray.ai is a composable AI integration and automation platform, and the Merlin release was aimed at a real adoption problem. The company's own release put it bluntly: agent experiences felt "clunky and disconnected" because agents lost context and forced users to start from scratch in every session. Memory, in that framing, is the cure for the amnesiac chatbot. An IT help desk agent that answers follow-ups, handles escalations, and remembers the user's earlier issue across sessions is a meaningfully better product. Merlin delivers that.

But a scheduled agent is not a chatbot with a calendar. Picture the Monday-morning pipeline digest. Every Monday at 7 AM, an agent on Tray.ai pulls the week's pipeline changes, summarizes what moved, flags deals at risk, and posts the digest to Slack. The team reads it, acts on it, and goes home. The following Monday, the agent runs again. Session memory does nothing for this job, because there is no session. There is a run. And each run starts with the same knowledge the platform ships, not with the thirty previous digests it already wrote.

What the Monday agent needs is a record of its own work: the deal it flagged as at-risk three weeks running, the rep's correction that the deal was fine and the real risk was elsewhere, the pattern it noticed about end-of-quarter discounting that nobody wrote down anywhere. That record compounds. Run one hundred is dramatically more valuable than run one, because run one hundred has a hundred weeks of context. Session memory was designed for a different problem: keeping a conversation coherent while it is happening.

There is an honest way to bridge the gap inside Tray.ai. Workflows can read and write to external stores, so teams build state by hand: decision logs, "already processed" tables, correction lists. Tray.ai's smart data sources help here too. The platform can sync structured and unstructured knowledge from file uploads or Google Drive with one click, vectorizing it automatically so agents stay grounded in current information. Between that and hand-built state tables, a careful operator can keep a scheduled agent reasonably informed.

The cost is that you become the memory engineer. Every table is a schema you designed, every write a convention you enforce, every retrieval a query you maintain. Stale entries pile up, two agents disagree about the same account, and the "already processed" list quietly becomes the most load-bearing database in the company. The platform handles the automation; the memory system is a side project that never ends.

And then there is the platform boundary. Tray.ai lets you build an agent once and deploy it anywhere: Slack, web apps, APIs, or running autonomously. That is real flexibility for deployment. It is not memory portability. The memory Merlin keeps stays inside Tray.ai. In practice, automation stacks are mixed. The pipeline digest runs on Tray.ai, the lead follow-up sequence runs on n8n, the enrichment step runs on Make. Each platform remembers its own slice in its own format, and none of them can read the others. The handoff between platforms is a memory reset, and most real workflows have several handoffs.

The fix is to stop asking each platform to remember and give the work itself a memory. A shared memory layer sits underneath the tools: the agent writes to it after every run, and every tool reads from it before the next one. Vilix AI is built for exactly this. It is cloud-hosted with zero infrastructure for you to manage, and the same memory is available to every connected tool over MCP, so what the Tray.ai agent learned on Monday is there for the n8n agent on Tuesday and the Claude session on Wednesday.

When you compare options, check three things. First, does it store full conversation history or only extracted facts? Facts tell you what was decided; history tells you why. Vilix AI keeps actual full conversations, and you can revisit the real conversation anytime. Second, what does it cost to try? The free plan is free forever, and the 7-day Pro trial asks for no credit card. Third, can you leave? Data is portable: export everything or delete it anytime. A memory layer with no exit is just a new silo wearing better branding.

So: Tray.ai remembers the conversation, and it does that job well. Your scheduled agent, though, does not have conversations. It has runs. Until those runs share a memory that outlives any one platform, every Monday starts at zero.

Get started at https://vilix.ai/?utm_source=vilix-blog&utm_medium=article&utm_campaign=tray-ai-remembers-the-conversation-scheduled-runs-start-blank