Perplexity adds memory-based personalisation to AI assistants

Perplexity has spent the past year pitching itself as the antidote to bloated chat interfaces and hard-to-trust AI answers.

Now it is taking a sharper swing at the incumbents with a set of personalisation features that transform its Comet Assistant and wider toolset into something closer to a persistent digital companion than a stateless chatbot.

The company’s new system allows its models to remember user preferences, interests and prior conversations.

Perplexity describes the upgrade as giving its assistant the ability to synthesise earlier interactions “like memory”, injecting them into future tasks so that responses become “smarter, faster, and more personalised, no matter how you work.” It is an explicit attempt to eliminate the slog of “context engineering”, where users repeatedly re-explain their goals just to keep an AI on track.

The underlying mechanism is designed to keep the remembered information tightly bounded. Rather than treating user history as training data, Perplexity retrieves context from a dedicated memory store and feeds it directly into the response workflow.

The company casts the distinction in simple terms: with this approach, “history is more than a dataset.” Memory is not folded into the model; it is supplied to the model.

That structure enables surprisingly granular use cases. The system can differentiate whether someone is injured or training for a marathon when recommending running shoes.

It can recall book suggestions for an upcoming trip or revisit earlier guidance on workplace issues. It holds onto structured preferences such as dietary needs, favourite brands and keywords, and a dynamic layer that pulls relevant details from recent exchanges.

Crucially, Perplexity stresses control. Users can switch memory off entirely, invoke incognito mode that disables both memory and search history, and rely on encryption for all stored information.

They can also opt out of contributing to model improvement through AI Data Retention settings. In other words, the assistant remembers only what it is explicitly allowed to remember.

The final twist is architectural. Perplexity says the personalisation layer works across every model it offers, from reasoning-heavy systems to faster specialised variants. Switching models no longer means rebuilding context from scratch. “Hours spent building unique memory are still valuable when a new model launches,” the company said.

In a moment when AI assistants feel increasingly interchangeable, Perplexity is betting that continuity, not novelty, will define the next phase of competition.

The recap

• Perplexity launches memory-based personalisation for its AI assistants.
• System stores structured user preferences and retrieves them as context.
• Memory works across all offered models, with user controls and encryption.