"Memory" has become the word for four unrelated mechanisms, which is why conversations about it go in circles. Someone means the context window. Someone else means the feature that recalls their coffee order. A third person means a document they uploaded last quarter. Deciding what your business needs is impossible until those are separated.
What is AI memory?
AI memory is the general term for information an AI system retains beyond a single turn — spanning the conversation window it can currently see, stored facts and preferences about a user, and knowledge it can retrieve again later.
The important structural fact underneath all of it: the model itself remembers nothing. A language model processes each request independently. Everything that feels like memory is stored outside the model by the surrounding product and re-supplied on the next request. When a system "forgets", it is almost never the model failing — it is the surrounding product deciding, correctly or otherwise, not to send something.
The four layers, separated
| Layer | What it holds | Who it belongs to | Lifespan | Freshness | What it cannot do |
|---|---|---|---|---|---|
| Working context | The current conversation the model can see | This conversation | Until the window fills or the chat ends | Exactly as current as what was said | Survive into a new chat |
| Per-user memory | Preferences and facts about one person | One user, in one product | Until edited or deleted | As current as the last update | Know anything the user never said |
| Stored knowledge | Documents, notes, uploaded files | Whoever can access the store | Until someone edits it | As of the last human edit | Update itself when reality changes |
| Retrieved knowledge | Passages found at question time | The corpus owner | Not retained after answering | As fresh as the index | Return facts nobody wrote down |
| Live business state | Current orders, spend, rankings, revenue | The business | Continuous | As current as the last sync | Exist without a connected system |
Most frustration comes from expecting one layer to do another layer's job. Asking per-user memory for this week's revenue fails not because memory is badly built but because revenue was never something the user typed into a chat.
Working context
The conversation window. It is genuinely capable — current Claude models support context windows up to a million tokens — but it is bounded and it is per-conversation.
Two behaviours matter. First, chat interfaces manage a full window on a rolling basis, so the earliest turns can drop out of a long conversation. Second, more is not better: Anthropic documents that as token count grows, accuracy and recall degrade, an effect they call context rot. A stuffed window is a worse window.
Per-user memory
The feature both ChatGPT and Claude now offer in various forms: the assistant retains selected facts about you across conversations, and you can inspect and delete them.
This is real and useful. It is also individual, product-bound, and sourced entirely from what you have said. It makes an assistant feel continuous. It does not make it informed about your company.
Stored knowledge
Documents, files, project instructions, wikis. The most under-rated layer, because a surprising amount of business context genuinely is stable prose: positioning, tone, policy, how you handle refunds.
Its single structural weakness is that it ages without telling anyone. A document is correct on the day it is written and silently drifts afterwards, and nothing in an answer built from it flags the drift. Knowledge base versus business context covers where that boundary sits in practice.
Retrieved knowledge
Retrieval-augmented generation searches an external corpus at question time and puts the retrieved passages into the model's context before it answers. Nothing is retained afterwards, which is a feature: the corpus is the source of truth and the index reflects it.
Retrieval scales stored knowledge past what fits in a window. It cannot invent facts that were never written down, which is precisely the category most operational questions fall into.
Live business state
The layer chat products do not provide, because it does not come from chatting. Today's orders. This month's ad spend against attributed revenue. Which queries gained impressions last week. Which products have not sold in ninety days.
This material only reaches an assistant through a connection to the system that owns it. There is no amount of remembering that substitutes for it.
AI memory vs persistent business context
Persistent business context is business information that exists independently of any conversation or assistant: maintained outside the model, refreshed from the systems that own it, and available to any client authorized to read it.
The distinction is about ownership and source, not storage duration.
Memory is user-shaped. It belongs to a person, inside a product, sourced from what that person said. If your colleague opens their own assistant, none of it is there.
Persistent business context is company-shaped. It belongs to the business, sourced from the systems of record, and any authorized assistant can read the same version. If your colleague connects a different AI product, the same facts are available.
A concrete pair:
An assistant remembering that you prefer weekly summaries on Monday mornings is memory.
An assistant knowing this week's Shopify revenue, current Google Ads spend, which campaigns are live, and which Search Console queries moved is business context.
The first improves how it talks to you. The second changes whether its answers are worth acting on. They are not substitutes, and a business that only invests in the first will keep wondering why the assistant is pleasant and useless.
Why "remember everything" is the wrong goal
The instinct is to retain more. Four reasons it backfires:
Old facts outrank new ones. A stored claim that margin is 40% does not know it has been superseded. Retention without an update path manufactures confident errors.
Every stored fact widens exposure. Memory is a disclosure surface. The more a single credential can recall, the more a compromised or over-scoped connection reveals.
Deletion becomes genuinely hard. If a customer asks to be forgotten, you have to know everywhere their details were absorbed. Context that is read from source systems on demand simplifies this considerably, because there is less copied material to hunt down.
Signal degrades. Retaining everything is the surest way to bury the fact that mattered.
The better goal is not maximum retention but correct retention: keep what is stable and hard to re-derive, read what changes, and label everything with when it was last true.
Freshness, provenance, permissions
Three properties decide whether any memory layer is safe to rely on.
Freshness. A stale answer is indistinguishable from a current one unless the system says otherwise. Any layer that can go out of date should carry an observation time.
Provenance. An answer you cannot trace is an answer you will re-check by hand, which cancels the time it saved. "Revenue is up 12%" is worth much less than "revenue is up 12%, from the orders synced at 06:00 today".
Permissions. Memory has to be scoped to a workspace, a brand, a customer. This is the failure mode that does not produce a bad answer — it produces a disclosure, which is worse and harder to notice.
What this means practically
- Want continuity with an assistant? Per-user memory and project instructions. Cheap, immediate, and no architecture required.
- Want it to know stable written facts? Documents or a knowledge base, with an owner and a review date.
- Want it to answer from numbers? Connect the system that holds them. Nothing else works, and no memory feature will close this gap.
- Want more than one assistant to know the same things? The context has to live outside any one product, which is what MCP exists to make possible.
Where Zylx sits, precisely
Zylx maintains persistent business context, not conversational memory. Concretely:
- A Business Brain is assembled from connected systems — Shopify, Stripe, Google Search Console, Google Analytics 4, Google Ads, Microsoft Clarity, Klaviyo, Ahrefs, Semrush, GitHub, and a crawl of the company's own site — and refreshed on a schedule rather than by hand.
- It also holds things a person deliberately teaches it: decisions, goals, constraints, and business facts that no connected system knows.
- Read tools return a freshness envelope with the data: where it came from, when it was observed, and whether it needs re-checking. That is provenance made visible rather than assumed.
- Access resolves to a workspace, expires, and is revocable.
What Zylx does not do: it does not store your chat history with Claude or ChatGPT, and it does not replace their per-user memory features. Those stay with those products. Zylx is the layer they read business facts from, which is a deliberately narrower job than "remembering everything".
For the wider design question this sits inside, see context engineering for business. For why context goes missing in the first place, see why AI loses business context.
Sources
- Anthropic: Context windows — how the window works, rolling management in chat interfaces, and context rot.
- Anthropic: Effective context engineering for AI agents — what makes up an agent's context.
- OpenAI: memory and controls in ChatGPT — current behaviour and user controls for ChatGPT memory.
- Model Context Protocol documentation.
- Zylx first-party: Business Brain sources, freshness envelopes, and access scoping in the Zylx Studio implementation, verified 2026-08-09.