Every conversation you have with a large language model has always ended the same way: it forgets you. Not because the model is bad at its job, but because statelessness is baked into how these systems were architected. Each session starts fresh, carrying no memory of what you told it yesterday, what preferences you expressed last quarter, or what your team is actually trying to accomplish. That constraint shaped how enterprises used AI tools, essentially as sophisticated search interfaces, not as ongoing collaborators.

That is changing. Four platforms are now building persistent memory into AI work assistants: OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini, and Amazon's Quick. The implementations differ significantly in scope, plan availability, and enterprise configurability. What they share is this: context about your employees is now accumulating on vendor infrastructure in ways that most enterprise IT teams have not governed.

Memory is not the same as context window

The two are often conflated, and the distinction matters for how enterprises should think about exposure. A context window is the amount of text a model can hold in active attention during a single session, measured in tokens. Claude currently offers a 200,000-token context window. Gemini reaches one million. These are impressive numbers, but they are session-scoped. When the session ends, so does the window.

Persistent memory is different. It operates outside the context window, storing synthesized facts and preferences across sessions in an external layer, then injecting relevant fragments back into the prompt when you return. The model doesn't re-read past conversations in full. It reads a curated summary of what the system decided was worth keeping. That curation step is the interesting part.

Technically, persistent memory systems combine a few components: a vector database that stores encoded representations of prior interactions, a retrieval mechanism that selects relevant memories based on semantic similarity to the current query, and an injection layer that prepends those memories to the active context at inference time. The result is a model that appears to know you, because contextually it does.

What the platforms are actually building

The implementations differ in ways that have real enterprise implications.

OpenAI was first with memory, and its system has iterated significantly. ChatGPT builds a persistent profile from your conversations through two parallel mechanisms: explicit memory, where you directly instruct it to remember something, and implicit memory, where it infers and saves facts automatically. The stored facts are unstructured plain text entries. Users with Plus or Pro subscriptions can now search and sort memories, but they cannot export them or transfer them to other platforms. There is a documented behavior worth noting for enterprise contexts: deleting a conversation does not delete the memories derived from it. You must manually remove them from a separate memory management page.

Anthropic launched automatic memory for all paid Claude plans, rolling out to Team and Enterprise first, then expanding to Max and Pro users. Its approach leans toward structured narrative organization, grouping stored information into categories rather than a flat list of entries. Users can view, edit, and delete individual memories, and can explicitly instruct Claude to remember something during a conversation. Claude also supports importing memory from other platforms. Extraction runs automatically in the background, but manual control over what persists is available, which gives Claude a transparency advantage over some competitors. Each project gets its own separate memory space, so context from one workstream does not contaminate another.

Google's Gemini sits in a more complicated position. A persistent memory feature exists, branded as Personal Context, but it is gated behind Gemini Advanced subscriptions and is not uniformly available across all account types. In enterprise Workspace deployments, administrators can disable cross-session memory retention entirely, and Google's own policies govern whether conversations are stored and for how long. What Gemini does carry that the others do not is authenticated access to Gmail and Google Drive for users who have connected those services. That integration is intentional by design and is why enterprise procurement teams should read the Gemini data terms carefully before enabling it at scale. The memory feature and the ecosystem access are separate controls, and most organizations have not separated them in policy.

Amazon Quick is the newest entrant and the one most worth watching for enterprise contexts specifically. Launched by Amazon Web Services in late 2025 and significantly expanded in April 2026, Quick is positioned from the ground up as an AI work assistant that connects to your apps, learns what matters to you, and takes action on your behalf. That phrase is not marketing copy. It describes the product's architecture: persistent memory of projects, files, communications, and work patterns, surfaced through integrations with Google Workspace, Microsoft Teams, Zoom, Salesforce, Dropbox, Airtable, and more. Where ChatGPT accumulates a profile from chat sessions, Quick accumulates a profile from your entire connected work environment. The data surface is an order of magnitude larger. The governance documentation to match it is not yet there.

On memory implementation maturity, my current ranking puts Claude first for combining automatic extraction with genuine user control: you can view, edit, delete, and explicitly instruct it to remember, with project-scoped separation that keeps workstreams clean. Amazon Quick ranks second for enterprise ambition and connector breadth, though its governance documentation has not yet caught up to its data surface. Grok ranks third as a capable but still consumer-oriented implementation with limited enterprise controls. Gemini ranks fourth given the plan-gating and the unresolved separation between memory controls and ecosystem data access. ChatGPT remains the most widely deployed and the most iterated, but memory locked inside a single platform with no export path is a structural liability as enterprise portability requirements harden. These rankings reflect current implementation, not long-term potential, and Amazon Quick could move quickly as its governance documentation matures.

Platform comparison: what each system remembers and who controls it

Why it is useful before why it is risky

The operational case for persistent memory is direct. An AI that knows your industry vertical, your preferred communication register, the platforms you run, and the context of ongoing projects produces materially better output from the first prompt of a session. You stop re-explaining yourself. The tool stops asking orientation questions.

For agentic AI workflows, memory is more than a convenience. An agent making decisions across a multi-step process needs access to prior decisions, established constraints, and organizational context that would otherwise have to be re-injected manually on every run. Without it, the agent reruns the same orientation work at the start of every session, burning tokens and compressing the window available for actual reasoning. Memory is the layer that makes accumulated work count.

Customer-facing applications gain the most visible payoff. An AI model that remembers a customer's prior service history, product configuration, and past preferences can reduce handle time and eliminate the re-authentication theater that makes customer service interactions frustrating. The payoff is real and measurable.

The governance problem nobody has solved

The risk is not theoretical. When an employee uses a ChatGPT Pro account for work tasks over several months, the platform builds a profile of that person's work patterns, vendor relationships, project priorities, and judgment calls. That profile lives on OpenAI's infrastructure. The enterprise IT team almost certainly does not know it exists, and almost certainly does not have a policy governing it.

This is the shadow AI problem applied to a new layer. Prior shadow AI concerns focused on data entered into prompts during a session. The memory problem is different because the data accumulates invisibly across sessions and persists after the session ends. It compounds over time.

Regulators are catching up faster than most enterprise AI programs have anticipated. The EU AI Act's general application provisions take effect in August 2026, covering high-risk systems in critical infrastructure, employment, and essential services. Transparency requirements under Article 50 already mandate disclosure when users interact with AI. A persistent memory profile built from work interactions sits squarely in the path of these requirements for many enterprise contexts.

There is also a procurement-level problem. As one enterprise security report noted, even when a vendor states that corporate data will not be used for model training, there is no reliable way to verify that claim. With memory features, the data surface is larger and the verification problem is harder.

Stale memories compound the risk. All three platforms have documented cases of outdated information persisting in memory and affecting response quality. In a personal productivity context, a stale memory about a restaurant preference is harmless. In a business context, a stale memory about a vendor relationship, a budget constraint, or a competitive position is an accuracy liability.

The portability question will matter more than the capability question

Right now, the conversation about LLM memory centers on what the model can do with stored context. The conversation that will matter in 12 to 24 months is what happens to that memory when you switch platforms, negotiate a new enterprise agreement, or the vendor changes its terms.

If an organization invests a year in building operational workflows around ChatGPT's memory layer, the organizational knowledge encoded in those profiles does not move with you. This is the current product reality, not a theoretical lock-in concern.

Gemini offers ZIP export. Claude supports memory import. Neither of those features is standard across the ecosystem, and the formats are not interoperable. Asking a vendor whether their memory feature supports export is now a reasonable procurement question, in the same category as asking about data residency or audit logs.

What a sensible enterprise posture looks like

The organizations handling this well are not the ones that blocked memory features wholesale. That approach fails because the productivity benefit is real and employees route around it. The organizations getting it right are treating memory as a data classification problem, not an AI policy problem.

That means categorizing what types of information employees are permitted to allow AI systems to retain, applying those categories to memory settings at the account or tenant level, auditing memory content periodically the way you would audit a shared drive, and including memory persistence, export, and deletion capabilities in AI vendor evaluation criteria.

It also means distinguishing between personal productivity use and enterprise workflow use. An analyst using Claude for research synthesis has a different risk profile than a customer success manager using it with customer data in scope. The memory feature should not be uniformly on or off across the organization. It should be configured to match the data sensitivity of the role.