My posts here so far have only been about the benchmark I'm working on. This time I want to introduce the company I work at, Wontopos.
Wontopos builds WOS: long-term memory for AI agents. It stores an end user's memories once, then recalls only the relevant ones per query so you can feed them into an LLM prompt.
It is not only for teams. If you are building a product, you use the SDK and your code decides exactly when to store and what to recall. If you just want a finished tool to remember (Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, Gemini CLI), you use MCP and write zero integration code. Same account underneath, so what one surface stores, the other recalls.
Each query returns a small, bounded set no matter how much is stored, so input cost does not grow with the store. A store that has been filling up for a year does not cost more to read from than a new one.
No LLM ever runs over your stored memories, on any model. Tablet runs no LLM at all. Scroll may use one to reformulate the query only, never your stored data.
No language is privileged: you can store and query in any language, mixed freely.
Four are live. Tablet is lean and returns less, Scroll returns more context. They read the same stored memories.
Tablet 1 (live) The original. About 1,200 tokens per query. Lean and fast, with no LLM anywhere in it.
Tablet 2 (live, and the default) About 1,000 tokens per query. It does three things Tablet 1 could not. The engine knows who said what, so you can register speakers and recall one person's words. It stores images as memories, and the caption may be empty, in which case the image is the memory and is searchable on its own. And it can be asked again: if one pass does not carry the answer, verify sends the engine back for memories it has not already returned, up to three times, with no LLM at any value.
Scroll 1 (live) About 3,700 tokens per query. Adds an LLM query-understanding layer over the same stored memories. That LLM reformulates the query only.
Scroll 1.2 (live, and the current Scroll)
Sentence-level recall over the same memories, so a fact said once, in passing, still surfaces. About 2,800 tokens per query, which is fewer than Scroll 1.
Book (in development, not priced) A fundamentally different design, not the context-for-accuracy trade. Built for one goal: never repeat the same mistake.
Every measurement we have published, on public benchmarks, with every run reported: