Retrieval-Augmented Browsing: How Agents Build Memory Across Sessions
A human who visits the same online store every week builds up knowledge about the site. They know where to find the sale section, how the checkout works, and which categories have the best selection. AI agents are starting to do the same thing.
Browsing With Memory
Traditional web agents are stateless. Each session starts fresh. The agent has no memory of previous visits, no knowledge of the site's layout, and no cached information about products or policies. It treats every visit like the first.
Retrieval-augmented browsing changes this. The agent maintains a persistent knowledge base, a collection of page summaries, extracted data, and structural observations from previous visits. Before interacting with a page, the agent queries this knowledge base for relevant information.
The result is an agent that gets faster and more effective with each visit. The first time it visits your site, it might spend 30 seconds figuring out the navigation structure. On the tenth visit, it already knows where everything is.
What Agents Store
The knowledge base typically contains several types of information.
Page summaries. Short descriptions of what each page contains. "The /delivery page explains shipping options: standard (3-5 days, free over £50) and express (next day, £7.99)." These summaries let the agent skip pages it has already read when the information has not changed.
Structural maps. The agent records how the site is organised. Navigation menus, category hierarchies, and the relationship between pages. This is similar to what agent memory and context systems track for repeat visits.
Extracted data. Specific facts pulled from pages: prices, product specifications, policy details, contact information. This data gets embedded as vectors for similarity search, as described in how embeddings work.
Interaction patterns. Notes about how the site behaves. "The search bar is in the top right. Autocomplete appears after 2 characters. The checkout requires an account." These patterns help the agent avoid trial and error on future visits.
The Retrieval Cycle
When an agent receives a new task involving your site, the cycle looks like this:
- The agent checks its knowledge base for information about the site.
- If relevant cached data exists, the agent uses it as context. It might already know the answer without visiting the site.
- If the cached data is stale or the question requires fresh information, the agent visits the site but uses its stored structural map to navigate directly to the right page.
- After the visit, the agent updates its knowledge base with any new or changed information.
This is conceptually similar to how retrieval-augmented generation (RAG) works in chatbot applications, but applied to live web browsing. The knowledge base acts as a long-term memory that supplements the agent's real-time observations.
Staleness and Accuracy
Here is where this gets interesting for site owners. Your content lives in an agent's memory for an indefinite period. If your prices change, your policies update, or your product catalogue shifts, the agent's cached version may be out of date.
An agent that cached your pricing page last month might tell a user that your Pro plan costs £29/month when you actually raised it to £39. That creates a poor experience for the user and a potential dispute for your business.
Structured data with timestamps helps. If your pages include dateModified in their Schema.org markup, agents can compare the stored version's date with the current page's date and decide whether to re-fetch. Without timestamps, the agent has to guess when to refresh.
HTTP cache headers (Last-Modified, ETag, Cache-Control) serve a similar function at the HTTP level. An agent that issues a conditional request with If-Modified-Since can quickly check whether a page has changed without downloading the full content.
The Compound Effect
As agents accumulate knowledge across multiple sites, they build a comparative database. An agent that has visited ten furniture stores remembers not just each store's products but how the stores compare. Price ranges, delivery options, return policies, customer review scores. All of this cached data informs future recommendations.
This means your site's information is not just competing in real-time searches. It is competing against stored memories from past visits to your competitors. If a competitor's cached data includes clear, structured pricing and your page was a confusing mess of conditional price tables, the agent's recommendation may favour the competitor even before visiting either site again.
What You Should Do
Keep your content accurate and up to date. This has always been good practice, but retrieval-augmented agents make it more consequential. Outdated information does not just sit on your site; it propagates into agent knowledge bases that influence future interactions.
Use structured data with timestamps. dateModified on Schema.org objects and proper HTTP cache headers give agents the signals they need to keep their cached versions current.
Maintain consistent URL structures. If the agent's stored structural map says your delivery information is at /delivery but you restructure and move it to /help/shipping, the cached map breaks. Redirects help, but stable URLs are better.
Make sure your site tells the same story in every format: HTML content, structured data, and meta descriptions should all be consistent. When agents cache your content, discrepancies between these sources create confusion in their knowledge bases.