A Smarter Tensor Memory Store for ChatVLM

ChatVLM has had memory for a while — but this release rebuilds it from the ground up. The Tensor Memory Store now organizes what it knows, recalls it more reliably, and gives you full control over every fact it keeps.

Topic documents

Instead of a flat pile of remembered facts, memories are now grouped into topic documents: living summaries organized around the things you actually talk about — a project, a codebase, a trip you're planning. When a new conversation touches a topic, the whole document comes along, not just fragments that happen to match keywords.

Recall you can see

Memory is only useful if it shows up at the right moment. We tuned recall thresholds against a new internal eval suite, and added observability so we can measure — not guess — when a remembered fact actually influenced an answer. The result is memory that surfaces more often when it should, and less often when it shouldn't.

Your memories, your call

The memory page has been redesigned around inspection and control:

• Browse and search everything ChatVLM remembers about you
• Add memories manually — tell it something once, explicitly, and it sticks
• Edit or delete anything, any time

Private by default in shared chats

One rule we consider non-negotiable: personal memories never surface in shared conversations. When you share a chat or collaborate in a workspace, your private context stays private — memory recall in shared contexts draws only on what belongs there.

All of this ships today for every plan, Free included. Open the memory page from your profile menu and take a look at what ChatVLM knows — and decide what it should keep knowing.