When a chatbot "remembers," one of two documented mechanisms is working: the context window — everything in the current conversation, including the system prompt, your messages, and the model's own replies, all present in working memory at once — or a memory feature, the vendor-documented capability where the assistant saves selected facts across sessions and re-injects them into later conversations, as ChatGPT's and Gemini's memory features and Claude's memory capabilities each describe in their own documentation. The distinction matters because the two mechanisms fail differently: context is bounded and forgets completely when the window closes; memory is curated, persistent, and can be wrong about you indefinitely unless you audit it.
RechargeMe publishes information, not advice. Behavior below follows the vendors' published feature documentation as of early 2026; settings names and defaults change, so verify in the product.
What exactly is a context window?
The model's working set, measured in tokens — the chunks models read text in. Everything the model currently knows about your conversation is inside it: instructions, all exchanged messages, pasted documents, tool results, and its own previous answers. Its documented consequences: very long conversations or pasted books consume it until earlier content falls out of attention, which users experience as the assistant "forgetting the beginning"; and the effective size varies by product tier and model version, per pricing and plan documentation. When a conversation ends, the window's contents are gone from the model's working set — nothing persists unless a memory feature saves it.
What do memory features document?
Vendor descriptions converge on a pattern: during conversations, the assistant extracts candidate facts — preferences, ongoing projects, biographical details you state — stores them, and uses them to personalize later conversations, with settings to view saved memories, delete individual items, and switch the feature off. ChatGPT's documentation describes both per-chat retained context and cross-conversation saved memories, plus controls excluding specific chats from remembering; Gemini's describes similar save-and-recall personalization tied to your account; Claude's describes memory capabilities that can persist information with user control, per Anthropic's feature pages. The common design point: memory is selective and editable, in both directions — you can delete what it saved, and it can save what you'd rather it hadn't.
| Dimension | Context window | Memory feature |
|---|---|---|
| Scope | This conversation only | Across sessions |
| Contents | Everything exchanged | Selected saved facts |
| Failure mode | Early content falls out | Wrong or stale facts persist |
| Control | Start a new chat | View, delete, disable |
Related stories: What actually counts as an AI agent? A definition with consequences · Training-data opt-outs: what vendors actually document you can control.
How do the two interact — and confuse?
The practical confusion: a chat that seems to remember may be doing either, and the difference is visible only in behavior. Within one long session, continuity is context — paste a document once and reference it for the rest of the session, and the model still sees it. A new session that greets you by your project's name is memory — a saved fact, re-injected. The documented trap is the seam: people paste sensitive context into a conversation, then start a "fresh chat" believing the tool has learned them, and are surprised both ways — that the fresh chat knows nothing (context gone) and that it still knows their preferences and job (memory persisting). Knowing which mechanism produced a behavior tells you which control fixes it: new chat for context, memory settings for memory.
What are the wellbeing and privacy angles?
Memory is the most personal dataset an AI product builds: it is, by design, a profile of you that changes model behavior. The documented controls — view, delete, disable — are the minimum habit to run once: audit what's saved, delete the stale and the sensitive, and decide whether the feature earns its convenience. Two structural cautions from the vendors' own policies: memories may be covered by different data terms than you assume (consumer training-use settings can apply), and shared or work accounts can leak personalization — the classic case of a family account whose assistant has merged everyone's stated preferences. For data-practice context generally, the US Federal Trade Commission's consumer guidance is the standing external reference.
How do you work with the limits instead of against them?
Documented vendor guidance and practice converge on a few habits. Keep important context explicit rather than implied — restate the load-bearing constraint (audience, format, definitions) when a conversation gets long, because context, not memory, is what the model is actually reading. Start a fresh chat when switching tasks, deliberately flushing context that would otherwise interfere. For recurring work, externalize memory yourself — a Project, a custom instruction, a saved prompt holds the stable facts under your control, rather than the assistant's curation. The model's attention is a rented workspace; your notes are owned property.
FAQ
- Why does my chatbot forget things mid-conversation? Long conversations exceed the context window and earlier content falls out of the model's attention — documented behavior, fixed by restating key facts or starting fresh with a summary.
- How do I stop my chatbot remembering me? Memory settings let you view and delete saved memories and turn the feature off — per each vendor's documentation; a new chat alone won't do it.
- Is memory the same as training on my data? No — memory is stored personalization for your account, separate from training-use settings, though both are governed by the vendor's current data policy. Check both.