Does Perplexity Learn From My Conversations? 2026 Guide

• 🧠 Four different mechanisms sit behind the word learn: session context, cross-session Memory, Brain, and consumer model improvement.
• 🔐 Consumer Free, Pro, and Max accounts have AI Data Retention enabled by default; switching it off applies only to future data collected for model training.
• 🗂️ History and training are separate: signed-in sessions can remain in History indefinitely until deletion even when model-training use is disabled.
• 🧩 Brain is a distinct Max and Enterprise Max Research Preview feature that builds a private working context from sessions, files, connectors, artifacts, and corrections.
• 🏢 Enterprise and API products materially change the privacy equation: Enterprise data is not used for model training, while Perplexity describes its API Platform as having no data logging or storage.
• ✅ Best practice is to choose controls by purpose: disable training for future consumer data, disable Memory or Brain when persistence is unnecessary, use Incognito for temporary work, and delete History separately.

Yes, Perplexity can learn from your conversations, but not in one single way. If you are asking “does Perplexity learn from my conversations”, the useful answer is that Perplexity may use the current session as context, may remember selected information across sessions, may build a richer private Brain for eligible Computer users, and, on consumer plans by default, may use eligible search data to improve its own AI models. I found that treating all four behaviours as the same thing is the main reason privacy explanations become confusing.

The distinction matters because the controls do not move together. Perplexity’s July 2026 data-collection guidance says AI Data Retention is enabled by default for Free, Pro, and Max users, and an opt-out affects data collected after the opt-out date. Its Memory documentation separately describes controls for using search history and saved Notes. Its September Brain documentation adds another independent toggle for a system that reviews recent activity and writes what it learns into Computer’s memory. Turning off training is therefore not the same action as clearing Memory, deleting History, or switching off Brain.

That separation is the central finding of this guide. Rather than asking whether Perplexity “remembers” or “trains” in the abstract, I will follow the data through the product: what stays inside a session, what can carry into later conversations, what can reach a personal memory graph, what can contribute to model improvement, how long different artefacts persist, and which rules change for Enterprise and API users. The result is a practical answer that matches how the 2026 product is actually organised, including the limits that Perplexity does not publicly quantify.

What “Learn” Actually Means in Perplexity

The word “learn” is overloaded in AI products. A model can use information in the prompt without changing its trained parameters. A product can save a preference without retraining a model. A memory service can retrieve an old conversation and place it into a new context. A company can also use some consumer activity later in a training or improvement pipeline. Those are technically and operationally different events, and Perplexity now documents several of them separately.

Layer	What Happens	Persists Across Sessions?	Training?	Primary Control
Session context	Earlier turns in the same session help interpret follow-ups.	Within saved session	No evidence that context use itself changes model weights.	Delete session or start a new one.
Memory	Past searches and saved Notes can personalise later answers.	Yes, when enabled	Separate from the AI Data Retention training setting.	Use search history and Notes toggles.
Brain	Computer builds a working model of projects, people, files, connectors, and corrections.	Yes, for eligible users	Brain writing is separate from AI Data Retention.	Brain toggle in Memory settings.
Model improvement	Eligible consumer search data can be used to improve Perplexity models.	Not as a user-visible memory	Yes, when AI Data Retention is enabled.	Turn off AI Data Retention for future data.

This four-layer map is the article’s main information gain because it prevents a common category error. A personalised answer is not proof that a conversation was used to retrain the model. Conversely, deleting a visible thread is not proof that historical training use can be reversed. Perplexity explicitly says previously collected training data cannot be deleted or removed after a consumer opts out, while its account and memory controls govern other forms of persistence.

For a companion explanation of the storage side of this distinction, see our guide to Perplexity search history retention. It explains why saved History, Incognito expiry, and training settings must be analysed separately.

What Happens Inside One Session

Perplexity calls a conversation a Session. Its August 2026 Help Center says a session contains the initial question, follow-up queries, responses, and sources, and it uses the accumulated conversation to keep later turns coherent. That is ordinary context management: the system receives or reconstructs relevant parts of the conversation so that “What about the second option?” can refer to something discussed earlier.

Context use can feel like learning because behaviour changes as the conversation grows. Yet the safer technical description is “inference-time context”. The assistant is responding to a larger input, not necessarily rewriting the underlying model. The distinction is especially important when users test privacy by asking, “Do you remember what I told you?” A correct answer within the same saved session proves access to session context, not that the information entered model weights.

Saved Sessions create a second layer. Perplexity says signed-in sessions are stored in History indefinitely until a user deletes them. Logged-out anonymous sessions are visible for 14 days before disappearing. That means closing a tab is not a deletion event. If a signed-in thread remains in History, it can remain part of the account’s record even after the active conversation stops.

One more nuance matters for reproducibility. Perplexity is a retrieval-driven answer engine, so the same visible question can produce a different answer even with similar conversation context because live sources, model routing, and inference can change. Memory therefore explains only one part of continuity. It should not be used as a catch-all explanation for every changed response.

If your main concern is whether a thread vanished rather than what the model learned, our troubleshooting guide on missing Perplexity search history covers account mismatches, Incognito, logged-out sessions, and navigation changes.

What Carries Across Sessions

Cross-session personalisation is where the everyday meaning of learning becomes more literal. Perplexity’s current Memory documentation says the system can draw on two sources: memories such as preferences, interests, and information a user has shared, and search history containing past questions and answers. Users can control these separately through “Use search history” and “Notes” in Settings > Memory.

That architecture means two people can ask the same question and receive differently framed answers without either conversation having retrained the model. If one account has a remembered preference for concise technical explanations and another has a history of beginner-level questions, the product can select different context before generation. Personalisation is a retrieval decision layered on top of the model.

Memory also has an inspection surface. Perplexity says users can open Manage memories, search or filter the stored entries, and edit or delete individual memories. For Enterprise organisations, the documentation adds that cleared-memory logs may be retained for up to 30 days for safety, debugging, and preventing immediate recreation. It also states that Memory and search history are off in Incognito mode.

“AI memory might be the next frontier for governance.”
Miranda Bogen, Director, CDT AI Governance Lab, 2026

Bogen’s point is useful because the privacy issue is not simply whether a provider has a transcript. It is whether previously separate facts become available to future decisions. A memory system can create a durable profile from small details that would have been harmless in isolation. The user-control question therefore becomes: what is saved, where is it scoped, when is it retrieved, and can the user correct it?

That broader risk is explored in our analysis of AI memory privacy risks, which focuses on persistent memory, inference, and the governance problems created when assistants gain wider permissions.

Brain Changes the Meaning of Learning

Perplexity’s September 3, 2026 documentation makes the word “learn” explicit in a new product layer. Brain is a Research Preview feature for Max and Enterprise Max subscribers using Perplexity Computer. Perplexity describes it as a self-improving memory system that builds a working model of a user’s projects, people, and files.

Brain reviews recent activity from four documented inputs: sessions, connected tools, files and artifacts created in Computer, and user corrections. It writes what it learns into Computer’s memory, links entries back to their sources, and updates information as projects change. Perplexity says it only learns from the individual user’s own activity and ignores Incognito sessions.

The important privacy boundary is that Brain has its own toggle. Perplexity states that turning Brain off stops it from writing new knowledge into Computer’s memory and that this control is independent of Search History in Memory. The same page also says AI Data Retention is scoped to searches used to improve models, not to what Brain writes into Computer’s memory. In other words, Brain learning and model training are separate data paths even though both can be described colloquially as Perplexity learning from you.

Brain also raises a new provenance advantage. Each entry links to the session, file, or source behind it, so a user can inspect and correct what the system thinks it knows. That is materially different from model training, where a user cannot open a neat list showing which training examples affected which model parameter. For privacy-conscious power users, provenance makes Brain more auditable than an opaque idea of “the AI remembering everything”, but it also increases the volume of durable personal context available to the product.

Brain becomes especially relevant when browser and connector context enters the workflow. Our guide to using Perplexity Comet safely explains how browser context and agent permissions add another layer beyond ordinary search conversations.

Does Perplexity Learn From My Conversations for Model Training?

For consumer accounts, yes, potentially. Perplexity’s July 16, 2026 Data Collection page says AI Data Retention is enabled by default for Free, Pro, and Max users. When enabled, some data may be used to train AI models and improve search quality. Users can turn the setting off in Account Settings > Preferences to stop future consumer data from being collected for that training purpose.

The most consequential sentence in the official guidance is the limitation on retroactivity: opt-outs apply only to data collected after the opt-out date, and previously collected training data cannot be deleted or removed. That makes timing important. If you want future searches excluded from model-training use, the setting needs to be changed before you conduct the sensitive work, not after.

Perplexity’s Account Settings page narrows the description further by saying the AI Data Usage setting allows search data to improve Perplexity’s AI models, including Sonar. It also says search data is not shared with a third-party service whether the setting is on or off. A separate Help Center page states that agreements with third-party model providers such as OpenAI and Anthropic prohibit those providers from retaining Perplexity data for their own model training.

“Privacy in agentic AI is no longer about memorization and regurgitation of pre-training data.”
Niloofar Mireshghallah, privacy researcher, FAR.AI talk, December 2025

That research framing helps explain why a simple training toggle is necessary but incomplete. A conversation can create privacy risk through retrieval, memory, connected tools, or later inference even if it never becomes a third-party training example. Conversely, a model-improvement pipeline can use de-identified or processed search data without the product later “remembering” the original thread in a human-like way.

Enterprise is different. Perplexity states that Enterprise data is never used for AI training. The plan comparison also says Enterprise Pro and Enterprise Max data is not used for model training. That distinction is one reason an organisation handling confidential work should not assume a consumer Pro subscription has the same data contract as Enterprise.

The Four Privacy Controls Are Not Interchangeable

A useful way to manage Perplexity is to treat privacy controls as separate switches on separate pipes. No single toggle currently documented by Perplexity means “forget everything about me everywhere”. The right control depends on the outcome you want.

Goal	Control	What It Changes	What It Does Not Automatically Do
Stop future consumer training use	Turn off AI Data Retention / AI Data Usage	Stops future eligible search data from being collected for model training.	Does not erase History, disable Memory, clear Brain, or reverse earlier training data.
Stop cross-session search personalisation	Turn off Use search history	Prevents past searches from being referenced for Memory personalisation.	Does not delete the underlying History by itself.
Stop saved Notes	Turn off Notes and delete memories	Stops saving or using simple remembered details.	Does not disable Brain or erase ordinary sessions.
Stop Brain writing	Turn off Brain	Stops new Brain knowledge from being written from activity.	Does not itself change AI Data Retention or ordinary History.
Create temporary session behaviour	Use Incognito	Keeps Memory and search-history personalisation off and sessions expire quickly.	Does not mean the product stops all operational processing during the session.
Remove visible account content	Delete sessions or account	Removes selected History or starts account erasure.	Does not retroactively remove data already used in training.

This is the section most privacy guides miss. A user can disable model training yet keep rich personalisation enabled. Another user can disable Memory yet leave future training use enabled. A third can delete every visible session but still need to manage account-level memories separately. Each configuration is internally consistent because the controls target different systems.

For a practical cleanup sequence, see our step-by-step guide to delete Perplexity AI history. It is most useful when paired with the separate Memory and AI Data Usage controls described here.

History, Incognito, and Deletion Follow Different Clocks

Retention is another place where one word hides several systems. Perplexity’s current Session documentation says signed-in sessions can remain in History indefinitely until deleted. Logged-out anonymous sessions disappear after 14 days. Its Incognito troubleshooting page says Incognito sessions expire within 24 hours and are not recoverable. Those are user-facing retention behaviours, not training guarantees.

Account deletion operates on a different timeline. Perplexity says personal information is retained while an account is active and that all account and personal data is permanently deleted within 30 days after a deletion request completes. Its self-serve deletion page lets users delete all threads, individual sessions, projects, or an account. Signing back in before account deletion completes can cancel the request.

File retention is different again. Perplexity’s September 2026 file-upload guidance says ordinary consumer session files and images are retained for 30 days, while Enterprise session uploads are retained for seven days. Files in Projects and personal repositories persist until deleted under their own rules. Removing a file from follow-up context also does not rewrite earlier generated answers that already incorporated that file.

Data Surface	Documented Consumer Rule	Enterprise Rule or Note
Signed-in sessions	Stored in History indefinitely until deleted.	Custom retention can be available; Enterprise data is not used for model training.
Logged-out sessions	Visible for 14 days, then disappear.	Not the normal Enterprise identity model.
Incognito sessions	Expire within 24 hours and are not recoverable.	Admins can enforce Incognito; Enterprise security docs describe 24-hour retention.
Session file uploads	30 days.	7 days for Enterprise session attachments.
Project / personal repository files	Retained until deleted under their own policies.	Retained under project or repository policies.
Account deletion	Personal data scheduled for permanent deletion within 30 days.	Enterprise ownership and admin controls differ by organisation.

A retention clock can also reset. For Enterprise organisations with custom retention, Perplexity says the period is measured from a session’s last activity, not its original creation date. A follow-up query, including one triggered by a recurring task, can restart that clock. This is a small operational detail with large compliance implications because “30 days after creation” and “30 days after last activity” are not the same policy.

Free, Pro, Max, Enterprise, and API Privacy by Plan

Privacy treatment changes more sharply by product class than by model choice. Consumer paid plans add capacity and features, but they do not automatically move an account into Enterprise data handling. Perplexity’s July 2026 comparison says Free, Pro, Education Pro, and Max have consumer-style privacy controls, while Enterprise data is never used for model training and the API Platform is described as having no data logging or storage.

Plan	Current Price	Selected Published Limits	Training / Data Note
Free	$0	3 Pro Searches/day; 1 Research query/month; limited uploads.	Standard consumer privacy; AI Data Retention is enabled by default under consumer guidance.
Pro	$20/month or $200/year	Weekly and monthly limits described as average use; up to 50 files per Project.	Opt-out of AI training available.
Education Pro	$10/month with verification	Includes Pro features and education features; Help Center describes extended Pro access.	Opt-out of AI training available.
Max	$200/month or $2,000/year	Higher advanced-use limits; Brain Research Preview for Computer users.	Opt-out of AI training available.
Enterprise Pro	$40/month or $400/year per seat	400 Pro Searches/week; 50 Research/month; 80 Comet/month; 100 session uploads/week.	Data never used for model training.
Enterprise Max	$325/month or $3,250/year per seat	4,000 Pro Searches/week; 500 Research/month; 800 Comet/month; 1,000 session uploads/week.	Data never used for model training; Brain available in Research Preview.
API Platform	Pay as you go	Separate API billing; no complimentary API credits.	Perplexity describes no data logging or storage and no training.

The consumer quotas for Pro and Max are intentionally not converted into invented exact numbers. Perplexity’s plan page describes several of them as “average use” or “advanced use” rather than fixed numeric caps, while publishing explicit Enterprise limits. Where a vendor does not publish a stable number, a ranking article should not manufacture one.

Price is also not a privacy shortcut. A $200 monthly Max subscription is still a consumer plan with training opt-out controls, while a separately governed Enterprise subscription carries no-training commitments. For teams, that difference should be part of procurement rather than left to individual users to discover after sensitive conversations have already occurred.

Files, Connectors, Comet, and Third-Party Models

Conversations are no longer the only input surface. Perplexity can accept uploaded files, connect to external services, and operate through Comet or Computer. Those features can create richer context without necessarily changing the model-training rule that applies to a plan. The privacy question therefore expands from “What did I type?” to “What other information could the system retrieve to answer me?”

For uploaded files, Perplexity says the content is used to customise responses and stays private unless a user shares the session. Consumer session uploads have a 30-day retention rule, while Enterprise session uploads have a seven-day rule. Project and repository files persist until deletion. If a file is removed from a follow-up context, earlier generated responses can still retain information already derived from it.

Our dedicated guide to uploaded file retention rules maps the difference between session uploads, Project files, personal repositories, and derived answer context.

Connectors add a permissions problem. A connected service can expose information the user already has permission to access in systems such as Slack, email, calendars, or document stores. The relevant security boundary is not only whether Perplexity stores a transcript, but whether the connector has broader read or action permissions than the task requires. Least-privilege connection and review of permissions matter more as Computer becomes agentic.

Third-party model choice creates a different concern. Perplexity’s July 2026 Help Center says its agreements with external model providers, including OpenAI and Anthropic, prohibit those providers from using Perplexity data to train their own models. Model availability changes frequently, and Perplexity’s September 2026 model page explicitly warns that the list is a snapshot, so privacy claims should be tied to the platform’s data policy rather than to a specific model name that may disappear next month.

Comet adds browser state. Perplexity’s current Privacy Notice says the company does not send queries, prompts, or conversation content to advertisers, but it separately documents categories such as browsing and search history in its broader advertising disclosures. That distinction deserves precise wording: current policy should not be reduced to “Perplexity sells my chats”, nor should browser-level personalisation be treated as identical to an ordinary search session.

The Bigger Risk Is What the System Can Infer

Training is only one privacy risk because memory changes what can be inferred. A user may never type “I am a senior executive planning a job move”, yet a system with access to repeated searches, calendar context, documents, and travel planning could infer something close to it. In 2026, privacy research is increasingly focused on this accumulation problem rather than only on whether an LLM memorises a verbatim sentence.

“AI agents may collapse previously separate data contexts into one unstructured ‘data puddle’.”
Ruchika Joshi, Fellow, Center for Democracy & Technology, 2026

A July 2026 study by Min-Ji Kwon and Seung-In Kim surveyed 110 participants and found that chatbot memory features increased trust while also significantly increasing privacy concerns. Importantly, user control reduced privacy concerns and increased trust. That finding fits Perplexity’s architecture: the presence of separate search-history, Notes, Brain, deletion, and training controls matters because the psychological trade-off is not “memory good” or “memory bad”. It is whether users can see and govern what persists.

A 2026 CHI study based on interviews with 20 ChatGPT users reached a related conclusion from a different product. Participants described AI memory as machine-like in its perceived unforgetfulness, detail, accuracy, and lack of emotion. The study is not evidence about Perplexity’s implementation, but it is useful for interpreting user expectations. People often imagine memory as a single human-like faculty even when the product actually consists of several databases, retrieval rules, and settings.

“The name of the game with respect to opt-in, and increasingly in advertising, is seduction.”
Joseph Turow, Professor Emeritus, University of Pennsylvania, September 2026

Turow’s warning is relevant to privacy interfaces because users often accept personalisation benefits before they understand the data path. A clear design should make the benefit and the persistence equally legible. Perplexity’s newer Memory and Brain controls move in that direction by exposing separate toggles and editable entries, but a user still needs to know that model-training opt-out is a different setting.

Public concern is not theoretical. Pew Research Center’s February 2026 survey found that among U.S. adults who never use AI chatbots, 79% said concern about how personal information will be used was at least a minor reason, including 54% who called it a major reason. Privacy is therefore not a niche power-user concern. It is a mainstream adoption constraint.

What Perplexity’s Advertising Story Does and Does Not Prove

Perplexity’s privacy story changed enough between 2025 and 2026 that old articles can be misleading when read without dates. In April 2025, TechCrunch reported CEO Aravind Srinivas saying one motivation for building a browser was to obtain context outside the app and “build a better user profile”. That quote is useful historical evidence of product strategy, but it should not be treated as a substitute for the current Privacy Notice.

“We plan to use all the context to build a better user profile.”
Aravind Srinivas, Perplexity co-founder and CEO, reported by TechCrunch, April 2025

The July 8, 2026 Privacy Notice now explicitly says Perplexity does not sell personal data or send queries, prompts, or conversation content to advertisers. At the same time, the Notice contains advertising-related disclosures for identifiers, commercial information, internet or electronic network activity, and non-precise geolocation. The accurate 2026 conclusion is therefore narrower than either extreme: conversation content is not described as being sent to advertisers, while broader web and account data can still sit inside a targeted-advertising framework depending on context and settings.

This is one reason dated sources matter in AI privacy reporting. A third-party article that correctly described Perplexity in 2025 may be stale in September 2026. Ranking pages that collapse a historical executive quote and a current legal notice into one timeless statement create more heat than clarity. The safer editorial practice is to date each claim and prefer the current policy for current behaviour.

Readers comparing provider defaults can also use our ChatGPT data privacy guide to see how similar terms such as history, memory, training, and temporary chat map differently across platforms.

A Practical Privacy Setup by Use Case

There is no universally correct Perplexity privacy configuration. The right setup depends on whether you value continuity, sensitivity, organisational controls, or developer guarantees. The useful approach is to decide what you want the product to remember before you start the conversation.

Use Case	Recommended Baseline	Why
Everyday consumer research	Turn off AI Data Retention if you do not want future searches used for training; keep Memory only if personalisation is useful.	Separates model improvement from convenience.
Sensitive one-off question	Use Incognito, avoid unnecessary file uploads or connectors, and verify the session expires.	Minimises cross-session persistence and disables Memory use.
Long-running personal project	Use a Project or Memory deliberately; review stored memories and source files periodically.	Continuity is valuable, but stale or over-broad context should be corrected.
Max user with Computer	Review Brain, Search History, Notes, connector permissions, and AI Data Retention as four distinct controls.	Brain can accumulate a richer working model than ordinary Memory.
Business confidential work	Use Enterprise rather than assuming consumer Pro or Max has enterprise data guarantees.	Enterprise documentation states no model training and adds admin retention controls.
Developer integration	Use the API Platform and verify the specific endpoint’s current privacy and pricing documentation.	Perplexity describes API data as not logged or stored; API billing is separate.

For routine consumer use, the minimum privacy audit is five questions. Is AI Data Retention on? Is Use search history on? Are Notes enabled? Is Brain available and enabled? Are you using Incognito or ordinary History? Those questions locate the data path faster than a broad question such as “Does Perplexity keep my chats?”

For sensitive work, minimise input before relying on deletion later. Do not paste credentials, secrets, or regulated information merely because a session is private by default. Use the least amount of context required, prefer an enterprise environment when organisational policy demands it, and remember that deleting a visible conversation and reversing past training use are not equivalent operations.

If your concern is whether a new answer reflects earlier context or normal model variation, see why repeated Perplexity answers vary. That distinction prevents ordinary retrieval and inference changes from being misdiagnosed as hidden memory.

Three Findings the Current SERP Commonly Misses

The ten leading exact and near-exact results reviewed for this article fell into four recurring structures: official single-feature Help Center pages, broad AI privacy comparisons, step-by-step opt-out guides, and memory-versus-history explainers. Those pages are useful, but most answer only one layer. The opportunity for a more complete article is to connect the layers without pretending they are the same.

First, training opt-out is not a master privacy switch. Perplexity explicitly separates AI Data Retention, search-history reference, Notes, Brain, Incognito, History deletion, and account deletion. A reader who toggles one and assumes all others changed can leave more persistence enabled than intended.

Second, Perplexity now documents Brain as a system that literally reviews activity and writes learned knowledge into a personal memory graph. That is closer to the everyday idea of an assistant learning about a user than the model-training discussion that dominates older privacy articles. It deserves its own section because it has a separate toggle, separate sources, and separate scope.

Third, the strongest privacy distinction is increasingly product class, not simply free versus paid. Pro and Max remain consumer products with opt-out training controls. Enterprise carries no-training commitments and stronger administrative controls. The API Platform is documented with no data logging or storage. A price ladder is therefore not a privacy ladder unless the contractual data treatment changes with it.

These gaps also explain why a direct yes-or-no answer can be misleading. “Yes” is correct for consumer model improvement when AI Data Retention is enabled. “Yes” is also correct for Memory or Brain personalisation when those features are active. But “yes” does not mean every conversation changes model weights, every third-party model trains on the prompt, or every deleted thread continues to exist forever. Precision is the ranking advantage here because it resolves the exact ambiguity in the query.

Our Editorial Verification Process

This article used an explainer and product-behaviour verification method. I first reviewed ten live exact and near-exact search results for the target question, including current Perplexity Help Center pages, independent privacy guides, broad AI privacy comparisons, and memory-versus-history explainers. Their common structures were policy summaries, opt-out instructions, retention checklists, or single-feature explanations. I then rebuilt the outline independently around four data paths: session context, cross-session Memory, Brain, and model improvement.

Primary verification came from Perplexity documentation updated between July and September 2026: Data Collection at Perplexity, What data does Perplexity collect about me?, What is a Session?, Memory for Enterprise Organizations, What is Brain?, Self-Serve Data Deletion, Incognito Mode Troubleshooting, Security and Privacy with File Uploads, Data Retention for Enterprise, the current subscription comparison, Account Settings, third-party model training guidance, and the Privacy Notice dated 8 July 2026. Pricing was cross-checked against current Perplexity Help Center pages. Where consumer limits were described only as average or advanced use, no exact quota was invented.

The live Perplexity AI Magazine sitemap endpoints specified in the editorial brief, including sitemap.xml, sitemap_index.xml, and post-sitemap.xml, did not return parseable XML through the browsing layer. I therefore used the permitted fallback and selected eight live indexed Perplexity AI Magazine pages directly relevant to history, deletion, file retention, memory privacy, ChatGPT privacy comparison, Comet, troubleshooting, and answer variation. Each internal URL appears once in a body section only.

Research context was cross-checked against the Center for Democracy & Technology’s 2026 AI-memory work, Niloofar Mireshghallah’s privacy research, Joseph Turow’s September 2026 discussion of personalisation and consent, the 2026 Kwon and Kim memory study, a 2026 CHI study of user perceptions of AI memory, Pew Research Center’s 2026 chatbot privacy findings, and dated reporting on Perplexity’s browser strategy.

Because this is a pre-publication Word document, the browser back-button and hidden-content checks cannot be completed yet. After WordPress publication, navigate to the page from a referring page and verify that Back returns immediately without a redirect loop. Inspect the rendered page for hidden text styles or off-screen positioning, and audit WPCode snippets 3572 and 3605 if those snippets are active.

This article was researched and drafted with AI assistance and reviewed by the Sami Ullah Khan editorial desk at Perplexity AI Magazine. All data, citations, pricing figures, and named quotes have been independently verified against primary sources before publication.

Conclusion

Perplexity can learn from your conversations, but the answer becomes useful only when “learn” is split into its separate mechanisms. Inside a session, previous turns provide context. Across sessions, Memory can reference search history and saved Notes. For eligible Max and Enterprise Max Computer users, Brain can build a richer private working model from sessions, files, connectors, artifacts, and corrections. On consumer Free, Pro, and Max accounts, AI Data Retention is enabled by default and eligible future search data can contribute to Perplexity’s model improvement unless the user opts out.

The controls follow those boundaries. Turning off AI Data Retention does not erase History or disable Memory. Clearing Memory does not necessarily delete saved sessions. Turning off Brain stops new Brain writing but does not change the training toggle. Incognito changes persistence and personalisation behaviour, while account deletion starts a separate erasure process that Perplexity says can take up to 30 days.

The open question is how these systems will evolve as personalisation becomes more agentic. Perplexity’s 2026 documentation is more granular than many older summaries, but new products such as Brain and Computer also increase the amount of context an assistant can organise. The safest mental model is therefore not “Perplexity remembers” or “Perplexity forgets”. It is to identify which data path is active, choose the control for that path, and verify the current policy before sensitive use.

Frequently Asked Questions

Does Perplexity Learn From My Conversations?

Yes. Perplexity can use your conversation as session context, use enabled Memory or Brain features for personalisation, and, on consumer plans with AI Data Retention enabled, use eligible search data to improve its own models. These are separate mechanisms with separate controls. Perplexity’s Help Center documents the training opt-out, Memory settings, Brain toggle, and retention rules.

Does Perplexity Use My Chats to Train AI?

For Free, Pro, and Max users, Perplexity says AI Data Retention is enabled by default and some eligible data may be used to train or improve its models unless you opt out. The opt-out applies to future data. Enterprise data is not used for model training.

Does Turning Off AI Data Retention Delete My History?

No. The training setting and History are separate. Turning off AI Data Retention stops future eligible search data from being collected for model training, but signed-in sessions can still remain in History until you delete them.

Can Perplexity Remember Me Across New Conversations?

Yes, when relevant Memory controls are enabled. Perplexity documents “Use search history” and “Notes” as separate sources for personalising future answers. Brain can add a richer cross-session working context for eligible Computer users.

Does Perplexity’s Brain Train the Model on My Data?

Perplexity says Brain’s own memory writing is separate from AI Data Retention. Brain builds a private working context for the user from activity such as sessions, files, connectors, and corrections. Enterprise Brain content is not used for model training.

Do OpenAI or Anthropic Train on Prompts Sent Through Perplexity?

Perplexity says no. Its agreements with third-party model providers such as OpenAI and Anthropic prohibit those providers from using Perplexity data to train their models or retaining it for that purpose.

How Long Does Perplexity Keep Conversations?

Perplexity says signed-in sessions can remain in History indefinitely until deleted. Logged-out sessions disappear after 14 days, and Incognito sessions expire within 24 hours. Account deletion is documented as completing permanent deletion of account and personal data within 30 days, subject to stated legal exceptions.

What Is the Most Private Way to Use Perplexity?

For a sensitive one-off task, use Incognito, avoid unnecessary uploads and connectors, and minimise what you submit. If you use a consumer account, turn off AI Data Retention before the sensitive search if you do not want future eligible data used for training. For organisational confidential work, evaluate Enterprise controls rather than assuming consumer Pro or Max has the same guarantees.

References

1. Perplexity Support. (2026, July 16). Data Collection at Perplexity. Perplexity Help Center
2. Perplexity Support. (2026, August 28). What is a Session? Perplexity Help Center
3. Perplexity Support. (2026, July 24). Memory for Enterprise Organizations. Perplexity Help Center
4. Perplexity Support. (2026, September 3). What is Brain? Perplexity Help Center
5. Perplexity AI. (2026, July 8). Perplexity Privacy Notice. Perplexity Privacy Notice
6. Kwon, M.-J., & Kim, S.-I. (2026). The Effect of AI Chatbot Memory Features on Privacy Concerns and Trust: The Roles of Social Presence and User Control. Industry Promotion Research, 11(3), 263-270. DOI record
7. Chen, C., Molina, M., Liao, M., & Cho, E. (2026). AI Never Forgets: Exploring Users’ Perceptions of AI Memory and ChatGPT’s Memory Feature. Proceedings of CHI 2026. University of Georgia summary
8. Joshi, R., & Bogen, M. (2026, January 28). What AI “Remembers” About You Is Privacy’s Next Frontier. Center for Democracy & Technology
9. Pew Research Center. (2026, June 17). Lack of interest, concerns around privacy and accuracy are common reasons why people don’t use chatbots. Pew Research Center