Smart Memory

Once you’ve saved a few hundred memories, loading all of them into every chat starts to hurt: wasted tokens, a crowded context window, and answers that get worse because the model is wading through things it doesn’t need. Smart Memory groups the memories in a bucket into categories, each with an AI-written summary, so your AI can load only the categories relevant to a conversation instead of the whole bucket. Smart Memory is available on all paid plans, including Core. In the dashboard the panel is called Smart Categories; the browser extension calls the in-chat toggle Smart Mode. They are the same feature.

​

Why it exists

Without categories, MemoryPlugin loads a bucket as one flat list. That works, but it has costs as the bucket grows:

​

How it works

Smart Memory works one bucket at a time:

1. Analysis: the AI reads the memories in a bucket to find themes.
2. Categorization: it groups related memories into categories such as “Work Projects” or “Health & Fitness”.
3. Summarization: it writes a short summary and additional context for each category.
4. Selective loading: at recall time, your AI loads the category summaries first, then pulls the full memories only from the categories it needs.

That last step is where the token savings come from. Think of a filing cabinet with labeled drawers: your AI reads the labels first, then opens only the drawer the conversation needs instead of tipping the whole cabinet onto the desk. Marketing describes it as “90% less tokens.” The mechanism is straightforward: loading a handful of category summaries costs far fewer tokens than loading every memory in the bucket, and the AI only expands a category when the conversation calls for it. The exact saving depends on how large your bucket is and how much of it a given conversation actually needs.

​

Requirements and limits

• Memory buckets only. File buckets and buckets shared to you cannot be categorized.
• At least 30 memories in the bucket before it can build categories.
• Won’t process a bucket over 600,000 tokens or 2,000 memories.

​

Building categories

​

What each category gives you

• Name: the theme the AI identified.
• Summary: a high-level overview of the memories in the category.
• Additional Info: deeper context about what the category covers. This is what helps your AI decide when to load the category, so it tends to matter more than the summary for recall quality.
• Memory count: how many memories are grouped there.

Each category also has a copyable category ID and a Filter by this option to narrow the memory list in the dashboard.

​

Resetting categories

To clear categories and start over, click Remove All Categories. The Reset categories for dialog asks you to type reset my categories to confirm, because it cannot be undone. Resetting removes every category, detaches category assignments from every memory in the bucket, and clears Smart Memory progress so you can recategorize from scratch.

​

Where categories are available

Once a bucket has Smart Categories:

• Browser extension: turn on Smart Mode and your AI can load specific categories in chat on ChatGPT, Claude, and other supported platforms.
• MCP server: MCP clients can call list_bucket_categories and list_category_memories to browse categories and pull memories from one.
• TypingMind plugin: exposes the same list_bucket_categories and list_category_memories functions.
• Custom GPT: does not expose the category tools today. It works with buckets and memories, but not category-level loading.

​

When it helps most

Smart Memory earns its keep on large buckets. If you have hundreds or thousands of memories, or several dense projects in one bucket, categories keep recall focused and cut token cost. A small bucket of a few dozen memories is usually fine to load whole, which is why the 30-memory minimum exists.

​

Next steps