Long-term memory
for AI agents,
in every language.
Lean and fast: about 1,200 tokens of context per query in ~320 ms, and 85.2% on LongMemEval-S. No keyword matching, no LLM in the retrieval path.
Finer recall, higher accuracy. Scroll 1.2 finds memory by the sentence, so a fact said once, in passing, still surfaces. It scores 92.3% on LongMemEval-S, up from Scroll 1's 90.7%, in about 2,800 tokens of context per query.
Learns from its mistakes. A fundamentally different design, not the context-for-accuracy trade. Built for one goal: never repeat the same mistake. When it gets something wrong, it learns so it won't miss that way again.
Codex — model pagecurl -X POST api.wontopos.com/api/v1/memory/recall
Don't take our word for it.
Ask the engine yourself.
This guide runs on your own LLM key, not ours. WOS loads the persona and everything it knows about the product and this page as memory, and the model simply speaks from it. We are showing you the engine by using it.
Connect your LLM key to start
The demo runs on your own key. It stays in your browser and is never sent to our servers.
Connect your LLM key
No login needed. Your key stays in your own browser - never sent to our servers.
Security: your key is never sent to our servers - it is saved only in your own browser's memory (sessionStorage) and wiped when you close the tab. We recommend using a temporary, scoped key and deleting it right after testing.
Type any model your key supports - we keep the latest as a default.
Three distinct
models.
Tablet, Scroll, and Codex are separate models, not versions of one another; each is built for a different need, so you pick the one that fits.
Lean and fast. Tablet returns about 1,200 tokens of context per query in roughly 320 ms, and scores 85.2% on LongMemEval-S.
Tablet 1 — full reportFiner recall, higher accuracy. Scroll 1.2 finds memory by the sentence, so a fact said once still surfaces, scoring 92.3% on LongMemEval-S, up from Scroll 1's 90.7%, in about 2,800 tokens per query.
Scroll 1.2 — full reportLearns from its mistakes. Codex is built on a fundamentally different design from Tablet and Scroll, not the same context-for-accuracy trade. It's made for one goal: never to repeat the same mistake. When it gets something wrong, it learns so it won't miss that way again.
Actual usage may differ from these numbers.
Your memory is yours. We hold it, and we don't look inside without your permission.
Our people and our models don't read your memories unless you ask us to, nothing ever trains on your data (no LLM runs in the retrieval path), one call erases a single memory or an entire user without a trace (GDPR-ready), and your LLM keys are sent per request and never stored while every memory stays isolated per account and user.
One file your AI
can read instantly.
Download a compact, LLM-friendly spec of the whole API. Drop it into your IDE or any coding agent and start building - no docs spelunking.
# Wontopos (WOS) - Long-term memory for AI agents > Pure semantic retrieval. No LLM in the path. ## Auth - API key + a storage id (user_id) ## SDKs - in YOUR app · Python · TS · Rust ## MCP (beta) - in AI tools · Claude Code, Cursor ## Pricing Tablet $2/$3 · Scroll $4/$8 per 1M + endpoints, engrams, speakers, caching, legal ...