Perplexity brings cross-model memory to its AI assistants

Starting a fresh chat only to explain your dietary restrictions, project details, or favorite brands all over again has always been one of the biggest friction points in using AI.

Perplexity announced a major upgrade to its personalization engine: AI assistants with persistent memory. The feature enables Perplexity, including its Comet Assistant, to remember key details and preferences across different conversations, pre-loading relevant context so answers get smarter and faster without manual prompt engineering.

A dedicated memory store, not training data

Instead of treating past conversations as generic training data where the model simply guesses what you might want next based on broad statistical patterns, Perplexity stores structured preferences in a dedicated memory vault.

This includes things like dietary requirements, frequent topics, and preferred brands. When you prompt the system, it actively retrieves only the relevant snippets from that memory store and injects them straight into the answer-generation process.

That means asking for a running shoe recommendation will automatically factor in whether you previously mentioned recovering from an injury or training for a marathon. You can also ask follow-ups on advice it gave you last week without having to dig up the old thread.

Users retain full control over what is saved. The memory feature can be turned off entirely, is automatically disabled in incognito mode, and all data is encrypted. Users can also opt out of data retention for model training.

Memory that travels across models

Perhaps the most compelling detail of the rollout is context portability.

Unlike traditional chatbots where your history is locked into one company's specific proprietary model, Perplexity allows your memory to persist across every model available on the platform.

You can switch to a deep reasoning model for heavy technical analysis, toggle over to a lightweight model for quick queries, and then try a new model tomorrow, all while drawing from the same underlying personal profile.

Why it matters

For power users and researchers, this finally makes Perplexity feel less like a search engine and more like a true second brain. You stop spending time re-explaining who you are and start getting answers calibrated to your situation immediately. The only potential snag is context pollution: if your tastes change or you ask about a topic outside your usual routine, an overly eager memory store might skew results unless you actively manage your stored profile.

From an industry perspective, Perplexity just made a brilliant strategic play against walled-garden rivals. Platforms like ChatGPT keep your personalization locked to their models. By making the memory layer model-agnostic, Perplexity transforms itself into the master interface for all frontier models, ensuring that whenever a shiny new third-party model launches, you bring your accumulated history with you rather than starting from scratch.

We think decoupling the personalization layer from the underlying LLM is exactly how third-party AI aggregators will survive. It gives users a tangible reason to stay, regardless of who wins the frontier model benchmark race this week.