What Anthropic says, and what it does not
There is no official figure. Anthropic’s article on usage limits lists the factors that affect how fast you reach your limit, and memory is not one of them. Neither the memory article nor the chat search article states a token cost. Anyone who quotes you an exact number for “memory overhead” is guessing.
What is documented:
| Factor | Counts against usage? | Source |
|---|---|---|
| Message length | Yes | Usage limit best practices |
| File attachment size | Yes | Usage limit best practices |
| Current conversation length | Yes | Usage limit best practices |
| Tool use (Research, web search) | Yes | Usage limit best practices |
| Model and effort level | Yes | Usage limit best practices |
| Artifacts and multi-step tasks (code, files, browsing) | Yes | Usage limit best practices |
| Project content reused | Yes, but cached, so counts less | Usage limit best practices |
| Memory | Not stated | Memory article |
| Chat search | Not stated; it appears as a tool call | Chat search and memory article |
Reasoning from what is documented
Everything Claude takes into account in a chat has to be in its context, and context is what costs usage. So memory is not free in principle: when Claude uses something it remembers, that content is part of the conversation. But memory in Claude is stored as individual topics, not as a transcript of everything you said, and Anthropic does not quantify memory’s net effect on usage: it depends on the context retrieved and the repetition avoided.
Chat search is different. When Claude searches your past chats, the search appears in the conversation as a tool call, with citations to the source chats. Tool use is on Anthropic’s list of usage factors, and the found snippets become part of the chat. If you ask Claude to “look up what we decided last month” often, that is the part to watch, not memory itself.
Where usage really goes
- Long conversations. Every new message carries the conversation so far. Starting a new chat for a new task is the biggest saving most people can make.
- Big attachments, uploaded again and again. Put reference material in a Project instead: Project content is cached and counts less when reused.
- Agentic work. Cowork tasks, Research and multi-step work use more than quick questions.
- Model and effort. A larger model at higher effort uses your limit faster.
Check where you stand in Settings > Usage: it shows how much of the five-hour session limit you have used and when the weekly limit resets. Verified Ultra accounts do not have a five-hour session limit.
The misunderstanding: turn memory off to save tokens
Turning memory off is not what Anthropic’s usage guide suggests. Its best practices include using Claude’s memory and project summaries to build context across conversations, so you do not repeat the same information. Anthropic does not quantify memory’s net effect on usage; it depends on the context retrieved and the repetition avoided.
There are good reasons to turn memory off for a chat (client separation, sensitive work); see how to turn off, reset or delete Claude memory. Saving tokens is not one Anthropic gives.
What this means for a team
On Team plans, usage limits apply per member and cannot be shared or transferred. So the cost of missing context is paid per person: each colleague who explains the same client background to Claude spends their own limit on it. A shared Project helps, because its knowledge is cached and everyone works from the same files. Personal memory does not help a colleague at all. More in does Claude have memory and how to use Claude Projects effectively.
Where Bombos fits
The expensive part of context is not memory, it is everyone rebuilding it by hand. Bombos is a plugin that gives every AI tool your team uses, Claude, Claude Code, ChatGPT, Copilot, Cursor and your own agents, one shared memory that retrieves only what a task needs from team knowledge, learned rules and what already lives in mail, documents and repos, with every answer traceable to its source. One person can try it first for their own agents.
Related questions
Where can I see how much usage I have left?
On paid plans, go to Settings > Usage. It shows how much of the five-hour session limit you have used and when the weekly limit resets. Verified Ultra accounts do not have a five-hour session limit.
Do Projects use more tokens?
Project knowledge is part of the context, but Anthropic says Project content is cached and counts less against your limits when reused. The cache expires after a period of inactivity.
Does turning memory off save usage?
Anthropic does not say. If you turn it off you may spend more messages repeating context, which the usage guide counts against you.
Is memory counted on the API?
On the API you pay for the tokens you send. With the memory tool, whatever Claude reads from memory files comes back as tool results in the conversation, which are input tokens.
