Gemini Personal Context: What It Remembers and How to Control It

Short answer: personal context is the umbrella for how Gemini personalises answers to you, and Google’s support documentation breaks it into three distinct sources: the memory of your past Gemini chats, your content and activity in Google apps you connect, and your stated preferences about how Gemini should respond. Those are separate settings with different privacy implications, and treating them as one switch is the most common mistake. Google also notes these features are not available to everyone, and that they require a personal Google Account rather than a work, school or supervised one.

Three Sources, Not One Feature

The reason people find this confusing is that “personal context” sounds like a single memory. It is three different mechanisms that happen to share a settings page.

Past chat memory. Gemini can draw on what you discussed in earlier conversations, so you do not re-explain your situation each time. Google has also been renaming the “past chats” feature to “memories” in the app, which is the same underlying thing described more plainly.

Connected Google apps. This is the source with the largest effect and the biggest privacy question. When you connect apps, Gemini can use your content and activity in them to answer. Google’s framing is that by securely connecting to the apps you use every day, Gemini understands your personal context well enough that you are not starting from scratch each time.

Stated preferences. Instructions you write about how you want responses: tone, length, format, what to assume about your expertise. This is the most controllable of the three and the most underused.

The important consequence: the first and third are about what you told Gemini, and the second is about what Google already holds. Different people will draw the line in different places, and knowing which is which is the point of separating them.

What Each One Actually Changes

If you are trying to decide where to start, custom instructions give the largest improvement for the smallest exposure. Most people never write any, then conclude the assistant does not understand them.

Bringing context from another assistant

One of the more practical recent additions is switching support. Google rolled out import tools in Settings for consumer accounts that let you bring your personal context, preferences and full chat history from other AI apps into Gemini.

The mechanism is worth knowing because it is unusual. You select the import option in Settings, paste a suggested prompt into your other assistant, and copy the summary it generates back into Gemini. Separately, you can upload a ZIP export of your chat history from another provider.

That first method is a neat solution to a real problem: your accumulated context in another product is not exportable as structured data, so the workaround is to ask that assistant to summarise what it knows about you and hand the summary over.

It is also worth pausing on before you do it. A summary of everything an assistant has inferred about you, moved into a second company’s product, is a meaningful step. Reading the generated summary before pasting it is a good habit, and most people will find it more revealing than expected.

What to Know Before Turning Everything On

Availability is limited and account-dependent. Google states these features are not available to everyone, and that you must be signed in with a personal Google Account. Work, school and supervised accounts are excluded.

Connected apps is the consequential setting. Past chats covers what you said to Gemini. Connected apps covers your mail, files and calendar. Those are different categories of exposure and deserve separate decisions.

Personalisation can degrade answers. Context helps when it is relevant and hurts when it is not. An assistant that remembers a project you abandoned may keep steering toward it. If answers drift oddly, stale memory is a likely cause.

Memory is reviewable. These features come with management controls, and reviewing what has been saved periodically is worth more than configuring it once. Google maintains a privacy hub covering how Gemini Apps handle data.

Shared devices change the calculation. Personalisation assumes one person per account. On a shared machine, or with an account family members access, it behaves in ways nobody intended.

When Personalisation Goes Wrong

It is worth describing the failure modes concretely, because when personalisation degrades an answer most people blame the model rather than the context.

The abandoned project problem. You spent three weeks discussing a plan you then dropped. Months later, unrelated questions keep bending back toward it, because the memory has no way of knowing the plan is dead. You did not tell it, so it assumes relevance.

The wrong-expertise problem. Early on you asked a beginner question in a field outside your own. That inference sticks, and now you get patient explanations of things you know well. It is subtle enough that people often do not identify the cause.

The context-flooding problem. With several apps connected and long chat history available, an assistant can pull in more context than the question warranted. The answer becomes broadly about your situation rather than sharply about your question, which reads as vagueness.

The shared-account problem. Personalisation assumes one person. On an account others use, it produces answers shaped by someone else’s interests, which is confusing before it is alarming.

The common fix for all four is the same and unglamorous: review what has been saved and delete what is no longer true. Personalisation is not self-cleaning, and a memory store nobody prunes gets worse over time rather than better. Ten minutes every few months is enough.

A Decision Framework for Each Setting

1. Always write custom instructions. Highest benefit, lowest exposure, and almost nobody does it. Three sentences about your role, what you already know, and how you want answers formatted changes output quality immediately.
2. Turn on past chat memory if you use Gemini for continuing work. For ongoing projects it removes real friction. For occasional unrelated questions it adds little and carries the most drift risk.
3. Connect calendar before mail. Calendar is mostly logistics and produces a large share of the practical benefit. Mail is the most sensitive content most people hold, and it deserves a separate, deliberate decision.
4. Connect files only for a real use case. If you actually want Gemini working with your documents, connect them. If it is speculative, wait until you have the use case.
5. Set a review date. Three months out, check what has been saved and delete what is stale. Personalisation without pruning gets worse over time rather than better.

Common mistakes right now

• Treating the three sources as one switch, and either enabling or refusing all of them.
• Skipping custom instructions, which is the best value on the page.
• Never reviewing saved memories, so an old project keeps colouring answers.
• Importing context from another assistant without reading the summary first.
• Expecting personalisation on a work or school account, where these features do not apply.

Getting Custom Instructions Right

Since this is the setting worth using and the one most often left empty, it is worth being specific about what a good one contains.

Who you are, in work terms. Your role and domain. “I am a physiotherapist working with post-surgical patients” stops the assistant explaining anatomy to you and starts it discussing rehabilitation protocols.

What you already know. The single biggest quality change. Telling it you are comfortable with statistics, or that you have never written code, removes an enormous amount of misjudged explanation in both directions.

How you want answers shaped. Length, whether you want the conclusion first, whether you want options or a recommendation. “Give me a recommendation, then the reasoning, not a list of considerations” is a transformative instruction for anyone who finds assistants evasive.

What to avoid. Filler openings, restating the question, excessive hedging. Naming these explicitly works better than people expect.

What you are working on currently. One line, updated occasionally, does more than months of accumulated chat memory because it is deliberate rather than inferred.

Four or five sentences covering those points will improve every answer you get, and it takes about ten minutes once.

The Skill Underneath the Settings

Personalisation is really a specific case of a general skill: controlling what context an assistant has, so its output is useful rather than generically plausible.

That skill applies whether the context arrives through a memory feature, a connected app, a document you attach, or an instruction you write. Knowing what to supply, what to leave out, and how to tell when stale context is degrading an answer is what separates people who find these tools genuinely useful from people who find them impressive but vague.

It also transfers across products, which matters given how often features get renamed and reorganised. Learning it deliberately in a structured sequence beats reverse-engineering it from settings pages. If you want a structured route in, explore Coursiv AI lessons and check current plan details on the official site.

FAQ

Your Next Step

Before touching any connection settings, spend ten minutes writing custom instructions: your role, what you already know, how you want answers shaped, and what to stop doing. Then ask a question you have asked before and compare the answer. Most people find that single change does more for output quality than every memory and connection feature combined, and it exposes none of your data to anything.