AI Terms Explained: Models, Apps, Browsing, Memory and Context

Quick answer

An AI model generates responses; an AI app provides the interface, settings and tools around it. Context is information available for the current response, while persistent memory can bring information into later conversations. Browsing retrieves web information; training changes the model’s learned parameters. Uploading a file can make its contents available for a task without immediately retraining the model. These are separate capabilities—not interchangeable labels. (Microsoft LLM fundamentals)

Before trusting an AI answer, ask three questions: What information could it access? What did it actually use? What might remain afterward? Use the distinctions below to investigate each one.

Features and controls can differ by account, region, platform and workspace. Check the current documentation for the exact product you use.

1. Model versus app: the generator is not the whole product

A model is the trained computational system that processes input and generates output. A large language model, or LLM, works with language represented as tokens. Its learned numerical values—called weights or parameters—help determine its responses. (Microsoft LLM fundamentals)

An application is the product through which you use a model. It can combine a model with instructions, memory, external tools and services. A web-search feature, for example, is a capability provided around the model—not proof that the model’s training knowledge continuously updates. (Google AI glossary)

This distinction matters when comparing tools. Instead of asking only, “Which model is better?”, ask:

• Which model or mode is selected?

• Which tools are available in this conversation?

• Which account or workspace settings apply?

• Can I inspect the sources used?

For a fictional example, imagine Mira asks an AI app to summarise a reading-club announcement. She then asks it to find the club’s current public schedule. Treat these as different tasks: the first uses supplied text; the second needs an appropriate external source.

Practical check: Compare the complete setup, not just the model name.

2. Context window: the working space for a response

Context is the information supplied to a model for the current response. It can include instructions, conversation messages and additional material. The context window is the limit on how much the model can process, measured in tokens. Tokens are pieces of text, not necessarily whole words. (Microsoft LLM fundamentals)

Think of context as a working desk, not a permanent archive. A document stored somewhere in your account and information supplied for a particular response are different things: an app may retrieve relevant material rather than supply everything it stores.

Mira might write:

“For this draft, use the announcement below. Keep the date, venue and organiser’s name unchanged.”

That prompt makes the essential material explicit. It also gives Mira clear details to check afterward.

For long tasks, keep a short, reusable task brief:

• Goal: What should the output achieve?

• Source: Which document or passage should it use?

• Constraints: What must stay unchanged?

• Unknowns: What should it leave unanswered?

Common mistake: Treating a large context window as permanent memory. One concerns processing capacity; the other concerns information brought into later interactions. (Google AI glossary)

3. Persistent memory: information carried between conversations

Persistent memory is an app-level feature that can use relevant preferences or details across conversations. Do not assume it is a complete, exact transcript. Check how the chosen product records, retrieves and displays remembered information.

Suppose Mira writes:

“I prefer short explanations with examples.”

Within the current exchange, that is an instruction. For it to influence a later conversation, the app needs some way to carry it forward—such as memory, custom instructions or relevant conversation history. These mechanisms have different controls.

Do not use the chatbot’s reply, “I’ll remember that,” as your only privacy check. Inspect the product’s memory settings and documentation.

Also separate stopping future use from deleting stored information. Check the removal steps for memories, chats, files and connected sources rather than assuming that deleting one item clears every separate record.

Practical check: Review what is remembered, correct stale preferences and examine the documented removal steps. Avoid assuming that a single switch clears every copy.

4. Browsing versus training knowledge: looking something up is different from learning it

Training knowledge comes from patterns learned during model training. A knowledge cutoff marks a boundary in that training coverage; it does not guarantee that every earlier fact is known or correct. A model can also generate confident but inaccurate answers.

Browsing, or web search, retrieves outside information for a task. It can help answer current questions without making that information part of the model’s learned parameters. (Google AI glossary)

For Mira’s schedule question, a useful prompt would be:

“Find the reading club’s current schedule on its official website. Cite the page and distinguish its publication date from the event date. If you cannot access the schedule, say so.”

Then check:

1. Did the response show evidence of web search?

2. Does the citation open the relevant page?

3. Does that page support the specific claim?

4. Is the information current enough for the task?

Search results and citations can themselves be incomplete, outdated or incorrect. Open the source rather than treating the citation marker as a guarantee.

Common mistake: Assuming that every answer from a browsing-capable app used browsing. Check the particular response and its sources.

5. File access versus model learning: reading is not retraining

Uploading a file can let an app summarise, compare or analyse its contents. That is different from training, which adjusts learned model parameters. Supplying material during ordinary response generation does not directly make those parameter changes.

Retrieval means finding relevant information to use in an answer. Retrieval-augmented generation, often shortened to RAG, combines retrieved material with response generation. A document used this way need not become part of the model’s training knowledge. (Google AI glossary)

However, “not immediately retraining” does not mean “never stored or used for model improvement.” Providers have separate data-use rules. Check the exact consumer or business policy and any separate terms for feedback before sharing material.

File access also does not prove complete reading. A successful upload may still produce incomplete analysis when a file is too large, complex, image-heavy or poorly structured.

Practical checks:

• Upload only material you are authorised to share.

• Ask for a page, heading or row supporting each important answer.

• Compare quoted details with the original.

• If coverage is incomplete, narrow the request or split the file.

For exact spreadsheet values, prefer a structured data file over an image of a table.

6. A worked example: follow one fictional conversation

Here is an illustrative workflow, not a recorded product test.

Mira is drafting an invitation for a fictional reading club. Her short source note says:

Willow Reading Club
Meeting: 14 November 2026, 3 p.m.
Venue: Community Room B
Bring one favourite short story.

First request: use supplied context

“Using only this note, draft an invitation. Preserve the date, time and venue. Do not add booking details.”

The expected check is straightforward: compare those three details with the note. If the draft adds a registration link, delete it unless independently verified.

Second request: check an upload

Mira attaches the same note as a document:

“Quote the meeting line before drafting. If you cannot read the file, tell me.”

The quotation gives her something concrete to compare. It is a check on this file and response—not evidence that every page of a longer document would be processed correctly.

Third request: investigate current information

“Is there a newer public announcement? Search the organiser’s official page. If you cannot find one, keep the supplied details and flag them for confirmation.”

Here Mira should inspect any retrieved source before accepting a change. Browsing provides evidence to evaluate, not automatic authority.

Fourth request: consider future personalisation

“For future invitation drafts, I prefer a friendly tone.”

Before relying on that preference next time, Mira should inspect the app’s memory or instruction controls. Carrying a preference forward and retraining a model are different processes.

Try it yourself: Label each request as supplied context, file use, browsing or persistent personalisation. Then identify which result you would verify manually.

7. A checklist for access, retention and retrieval

Before using an unfamiliar AI tool, make a small capability record:

Question	What to inspect
What am I using?	App name, selected model or mode, account type
Can it browse?	Search control, response citations and source pages
Can it use files?	Attached files, supported formats and reading limitations
Can it reach connected services?	Connected-app permissions and workspace controls
Can it personalise later chats?	Memory, history and custom-instruction settings
What remains afterward?	Chat, file and memory retention documentation
Can content improve models?	Training/data-use settings and feedback exceptions

Do not collapse all privacy questions into “Is memory off?” For example, Google’s Gemini documentation says temporary chats and chats with Keep Activity off are retained with the account for 72 hours for specified service and protection purposes. It separately explains model-training and feedback rules. Retention and training are different questions. (Gemini Apps Privacy Hub)

If a capability is missing, check the actual account, platform and workspace documentation before concluding that the model cannot perform it. Availability can vary.

The most useful habit is to name the uncertainty precisely: “I don’t know whether this response used the web,” “I haven’t checked whether this file remains stored,” or “I need to review memory settings.” Each statement points to a practical next step.