Your Scheduled Research Agent Rediscovers the Internet Every Morning. Here's the Memory Pattern.

Your Scheduled Research Agent Rediscovers the Internet Every Morning. Here's the Memory Pattern.

Every Monday at 6 AM, a research agent wakes up and produces a market brief: funding rounds, product launches, pricing moves, the works. Every Monday the brief lands on time. And every Monday the agent builds it the same way: by rediscovering the entire internet from nothing.

It starts with the landscape. Which companies are in this space, what they charge, who funds them. It researched all of this last Monday, and the Monday before that, and the thirty Mondays before that. But between runs it kept nothing, so it re-reads the same forty pages, re-extracts the same facts, and re-writes the same background section before it gets to anything new. The new stuff, the actual reason the brief exists, gets whatever time and tokens are left.

Scheduled research agents fail in a specific way. They do not forget the task. They forget the context they already paid to build.

Count what re-reading costs you

Put rough numbers on it. Say the agent reads 40 pages a week and 35 of them have not changed since last week. A typical page costs a few thousand tokens to fetch and summarize. Multiply that by 35 unchanged pages, by 52 weeks. You are paying to re-learn your own market dozens of times a year, and the payment is not only money. It is the context window.

A long-context model can hold a lot, but every token spent re-reading an unchanged page is a token that is not spent on the three pages that moved. The brief's quality is bounded by how much of the window is left after the déjà vu. Agents that re-read everything produce briefs that read like everything is equally new, because to the agent, everything is.

There is also a quieter cost: the agent never develops taste. A human analyst who covers a beat for a year learns which sources break news and which ones repackage it, which blogger has real access and which one rewrites press releases. Your agent does that work every week and retains none of it. It is permanently a first-day analyst.

Three things the agent should carry between runs

Memory for a research agent is not one thing. It is three records that play different roles.

The reading record is the ledger: which pages the agent has seen, when, and what they contained when it saw them. This is what turns "research the market" into "research what changed." Before the agent fetches a URL, it checks the ledger. Same content as last time? Reuse the summary. Changed? Read and re-summarize. New? Read it fresh.

The conclusions record is the judgment layer. This is not what sources said; it is what the agent thinks about them. Which sources break stories first. Which ones are reliably late. Which topics are exhausted and which are heating up. Conclusions are the closest thing a research agent has to expertise, and they are pure profit: every run makes the next run's source selection smarter.

The open-questions record is the agenda. What the agent still does not know. "Is the rumored pricing change real?" "Who is the new VP of product at the competitor?" A stateless agent treats every run as a blank page, so it never maintains an agenda. An agent with an open-questions list wakes up with priorities instead of a to-do list that says "everything."

Wire the read-before-research loop

The pattern that makes this work is a loop with two mandatory steps.

Step one happens at the start of every run: the agent reads its memory before it touches the web. It pulls the reading record, the conclusions, and the open questions, and plans the run around them. The unchanged pages get skipped. The trusted sources get checked first. The open questions become the research agenda. This planning step is the whole trick. Memory that is never read before acting is not memory; it is an archive.

Step two happens at the end of every run: the agent writes back. New pages join the reading record. Verdicts get updated when a source surprises the agent, in either direction. Questions that got answered move off the open list; new ones the run surfaced get added. The write-back is what makes the system compound. Skip it and the memory goes stale within weeks, and a stale memory is worse than none because the agent trusts it.

Hosted memory vs. the DIY route

You can build the three records yourself. A JSON file in the agent's working directory gets you surprisingly far: readable, editable, free. It falls apart when the agent runs in containers, moves between machines, or when a second agent needs the same memory. Then you are looking at a database, which is durable and queryable and means you now run infrastructure so your agent can remember its reading list.

The alternative is a hosted memory layer over MCP. Vilix AI is cloud-hosted with zero infrastructure to manage, and the same memory follows the agent across every AI tool you connect. It keeps full conversation history rather than just extracted facts, so the agent's reasoning about a source is preserved along with its conclusions, and retrieval is semantic, so "the pricing page that changed" is findable by meaning, not just by URL. The free plan is free forever, the Pro trial runs 7 days with no credit card, and you can export or delete everything whenever you want. For a research agent, that means the reading record, the conclusions, and the open questions are simply there at the start of every run, on any machine the scheduler uses.

Either way, the requirement is the same: the memory has to be readable at the start of the run and writable at the end, and it has to survive the scheduler killing the process in between. If your setup does that, you have a research agent with a memory. If it does not, you have a very expensive way to re-read the internet.

The brief that builds on itself

Give the agent the loop and watch the briefs change. Month one, the background section is long because the agent is still building its records. Month three, the background shrinks to "no changes since last week" and the brief leads with what moved. Month six, the open-questions list is doing the job of an assignment editor, pointing each run at the frontier instead of the archive.

That is the actual product of a research agent with memory: not a better single brief, but a brief that gets better on its own. The agent stops rediscovering your market every Monday and starts covering it like someone who has been on the beat for a year. All it needed was a place to put what it learned, and a habit of reading it before it starts.