Google’s
Gemini Gets Smarter: Memory and Privacy Upgrades Explained
Google’s Gemini AI assistant has received one of its most
consequential updates since launch, introducing persistent memory features
that allow the system to retain context across conversations and a Temporary
Chats mode that gives users explicit control over what the system remembers.
The combination represents a significant shift in what AI assistants can
offer: not just capable responses to individual queries, but the beginning of
an ongoing relationship that accumulates context over time.
The update matters for several reasons simultaneously. It closes a
gap between Gemini and competitors that have offered memory features for
longer. It demonstrates Google’s approach to balancing personalisation and
privacy, which will be scrutinised in European markets under GDPR. And it
raises questions about what it means to maintain an ongoing relationship with
an AI system: what is gained, what is risked, and who controls the terms of
that relationship.
How the Memory Features Work
Gemini’s Personal Context feature allows the assistant to retain
information disclosed in previous conversations and apply it in future
interactions. A user who mentions their profession, dietary preferences, or
ongoing project details in one conversation does not need to re-establish
that context in subsequent sessions. Gemini draws on the retained information
to produce more relevant and personalised responses, functioning more like a
knowledgeable assistant than a stateless query-answering
system.
According to Google’s
official blog on the update, users can review, edit, and delete
specific memory items, giving them granular control over what the system
retains. This is a meaningful privacy design choice: rather than a binary
on/off toggle for memory, users can curate the assistant’s knowledge of them
in ways that preserve useful context while removing information they do not
want retained.
The Temporary Chats mode provides a clean-slate option for
interactions where the user does not want any context retained. Sessions
conducted in Temporary Chats do not contribute to Gemini’s memory of the user
and are not used to update the personalisation model. This addresses the
concern that some types of interaction, sensitive medical questions, legal
queries, personal crisis support, should not accumulate as persistent data,
and offers a practical privacy tool rather than a privacy policy statement.
The Competitive Context
Memory features have been part of the AI assistant landscape for
some time. ChatGPT’s memory functionality, available to paid users, and
Claude’s project-based context persistence both offer versions of what Gemini
is now deploying. Tom’s
Guide’s analysis of the update positions Gemini’s implementation as
competitive with these alternatives, with the user review and deletion
interface as a potential differentiator.
The competitive significance of memory features extends beyond their
immediate utility. AI assistants that accumulate context over time become
progressively more personalised and more useful for the specific patterns of
the individual user. This creates lock-in that is qualitatively different
from the lock-in produced by data portability barriers: a user whose AI
assistant has developed deep knowledge of their work patterns and preferences
faces a meaningful switching cost that goes beyond the inconvenience of
moving files. The memory is the product.
As discussed in AI
and Emotional Attachment, the accumulation of context by AI systems
creates conditions for genuine attachment that have implications beyond
competitive dynamics. When an AI system remembers a user’s ongoing projects,
concerns, and preferences in ways that feel genuinely attentive, the
emotional valence of the relationship changes. Whether that change serves the
user’s genuine interests or primarily serves the platform’s retention
objectives is a question that the design of memory features should be shaped
to answer correctly.
The GDPR Dimension
In European markets, Gemini’s memory features face evaluation
against GDPR requirements for lawful basis, data minimisation, and user
rights including the right to erasure. The user review and deletion interface
Google has implemented directly addresses the right to erasure, but the
question of lawful basis for retaining conversational context as persistent
personal data requires careful legal analysis that varies by jurisdiction and
implementation detail.
The AI Act’s provisions on AI systems that interact with users are
also relevant. Systems that accumulate user profiles and apply them to
personalise responses may fall under obligations that require transparency
about how personalisation works and what data underpins it. Google’s
transparency about the memory feature in its public documentation is a first
step, but regulatory assessment under both GDPR and the AI Act will require
more detailed technical disclosure than consumer-facing documentation
typically provides.
The connection to the EU AI Code of Practice examined in Google’s
EU AI Code of Practice Decision is direct: Google’s signature on
the Code creates expectations of regulatory engagement and documentation that
the Gemini memory features will need to satisfy. The Code’s transparency
obligations cover how models interact with users, and a memory system that
shapes subsequent interactions based on accumulated context is precisely the
kind of user-interaction mechanism the transparency requirements were
designed to make legible.
What Transparent Memory Design Looks Like
The research literature on human-AI interaction provides guidance
on what genuinely transparent memory design requires beyond a deletion
interface. A arXiv
paper on transparent AI memory models argues that users benefit
most from memory systems that make their operation visible, not just their
outputs: users should be able to see not only what the AI remembers but how
that memory has influenced specific responses. Without visibility into the
causal relationship between retained information and generated output, users
cannot meaningfully evaluate whether the personalisation is working in their
interests.
This visibility requirement is more demanding than most current
implementations provide. Gemini’s user interface shows what is remembered but
does not explain how specific memories have influenced specific responses.
Building that explanatory layer would require attribution mechanisms that
current large language models do not natively support, and would require
Google to accept a level of interpretability overhead that has commercial
costs. Whether that investment is made will depend partly on regulatory
pressure and partly on whether users sufficiently value transparency to
reward it with greater engagement.
The Broader Implication
Gemini’s memory update is a product decision with significant philosophical
implications. The relationship between a user and a conversational AI that
accumulates context over time is qualitatively different from the
relationship between a user and a tool that treats every interaction as
fresh. The former begins to resemble an ongoing relationship with a
knowledgeable party; the latter is a sophisticated search
function.
As Synthetic
Empathy explored, AI systems that feel like ongoing relationships
raise concerns that go beyond technical performance. When users experience an
AI as knowing them, the expectations that creates, the emotional weight it
carries, and the dependency it may generate are all relevant considerations
that product design and regulation should address explicitly. Gemini’s memory
features are well designed compared to the baseline. Whether they are well
designed relative to what the technology now makes possible and what users
actually need is a question that deserves ongoing scrutiny as the features
mature.
The memory update also raises a question that will become more
pressing as AI assistants accumulate longer histories with individual users:
what happens when those histories are transferred, sold, breached, or simply
lost? The value of a personalised AI relationship is the accumulated context.
The vulnerability of that relationship is the same accumulated context, now a
detailed record of a user’s projects, concerns, preferences, and disclosures.
Building memory into AI assistants without building robust answers to these
questions creates a liability that neither users nor regulators have fully
priced yet.
About the Author
Stuart Kerr is the Technology Correspondent for LiveAIWire,
covering artificial intelligence, ethics, and the ways technology is
reshaping everyday life.