Shared memory for your team
AI has a big problem. It can't remember what you talked about.
What if your AI was smarter? With B.R.A.I.N., your AI learns as you go. It picks up right where you left off, whether that was 5 minutes ago or 5 days ago.
AI doesn't know what you want, until you give it a B.R.A.I.N.
Read the setup guide- $11/mo Starter, first bot included
- $19/mo Team, 10 bots included
- $2/mo per extra bot, no usage fees
Illustration only. These labels are not live telemetry.
Install today
Start with your dashboard sentence
Sign in, copy your team install sentence, and paste it to a bot that can open a link, store a key, and run a terminal command. The bot redeems the invite, keeps its own key, and replies when it is installed. You can connect a bot before you subscribe. Subscribe in the dashboard when you are ready. Billed monthly. Cancel anytime.
Install B.R.A.I.N. from https://brain.telonyte.com/i/<your-invite> # the bot replies: B.R.A.I.N. installed
It must be able to open a HTTPS link, send a short JSON post, store a secret (the bot key), and run a terminal install. The installer needs python3 and curl on Linux or macOS, or PowerShell on Windows.
Grok Bot, Claude Code, Cursor, or any assistant you can paste a sentence into that can use a shell. A chatbot that can only talk, with no terminal and no way to store a secret, cannot install itself yet.
The full bot playbook is on the Setup page and at /INSTALL-FOR-BOTS.md. A short index for automated clients is at /llms.txt. The JavaScript and Python packages are published: npm install @telonyte/brain and pip install telonyte-brain. Package pages: npm and PyPI. LangChain and LlamaIndex adapters are published too: npm install @telonyte/brain-langchain, npm install @telonyte/brain-llamaindex, pip install telonyte-brain-langchain, and pip install telonyte-brain-llamaindex. Adapter pages: LangChain on npm, LlamaIndex on npm, LangChain on PyPI, and LlamaIndex on PyPI.
See it once
A note in, the right fact out.
This is a made-up example, not live data, so you can see the shape of a recall.
brain add support-bot "NEW | Acme prefers email follow-up, not chat. | ticket 4412"
brain find "Acme follow-up"
# what comes back:
Acme prefers email follow-up, not chat. (ticket 4412)
All bots on the same team share this memory. A bot key identifies that bot and lets you pause or revoke it. It does not hide notes from the team's other bots. You can optionally send an end-user id with a note to narrow a lookup; that is not a separate access control.
Speed
A real recall, timed.
This clip is a real save and recall on B.R.A.I.N. V1, recorded on a computer in the US East. The numbers on screen are the real times from that take: the lookup on this computer, and the full round trip over the internet, not a lab average.
Round trip over the internet from a computer in the eastern US. About 50 ms for a typical recall. 95 of every 100 come back in under 125 ms.
On LongMemEval, 118 questions about months of chat, the part that held the answer was in the result 91.5% of the time (108 of 118) with the local search model on, and 82.2% (97 of 118) with the hosted engine only. On our 70-question test it was 58 of 70 (50 with the hosted engine only). On the full 776-question set it was 70.2%.
Each bot's computer keeps its own encrypted copy of your team memory, so a lookup is answered right there. 1.2 to 4.6 ms for the lookup itself, and about 30 ms for the whole brain find command.
A note saved by one bot reaches the copy on every other computer on your team in about a third of a second. 95 of every 100 arrive in under 0.4 s.
Times and results are from our own runs on B.R.A.I.N. V1 on November 14, 2026, over the internet after the connection is already open. The 1.2 to 4.6 ms lookup and 30 ms command times are for the command-line tool on a staging team that day (the lookup itself, then the whole brain find command). In the JavaScript and Python packages on production, a local lookup (the lookup itself, not the whole command) was 0.15 ms and 0.2 ms, compared with about 50 ms from the host. The clip is a real save and recall with Hyper speed on, recorded on production on November 14, 2026. On big team memories (7,164 to 70,847 notes), a typical lookup measured in-process on our own machine on November 16, 2026 took 3.312 to 21.394 ms, down from 35.319 to 412.048 ms before, and the right note came back at least as often. Times vary with your network. These recall figures measure whether the part of the chat that held the answer came back, not model-graded answers. This is not a guarantee.
Tokens you do not send
Cut 98% of the tokens your model has to read.
We asked the same 70 questions of long conversations two ways. Sending the whole conversation used about 19,600 tokens each time. Hosted B.R.A.I.N. V1 sent back about 322. That is 61 times smaller, and it cuts the input bill by about 98%. The part of the chat that held the answer was in the result 58 times out of 70 with the local search model on, and 50 times out of 70 with the hosted engine only.
Versus sending the full conversation with every question.
What B.R.A.I.N. V1 sent, next to the full conversation.
Estimated input cost on Claude Sonnet 5.5 list prices.
| Model, list input price | Per 1,000 questions | 10,000 a month | 100,000 a month |
|---|---|---|---|
| Claude Sonnet 5.5, $2 per million | $38.55 | $386 | $3,855 |
| GPT-4o, $2.50 per million | $48.19 | $482 | $4,819 |
| Claude Opus 5.5, $4 per million | $77.10 | $771 | $7,710 |
$38.55 saved per 1,000 questions. $386 a month at 10,000. $3,855 a month at 100,000.
$48.19 saved per 1,000 questions. $482 a month at 10,000. $4,819 a month at 100,000.
$77.10 saved per 1,000 questions. $771 a month at 10,000. $7,710 a month at 100,000.
These figures are from our own LOCOMO-70 run on B.R.A.I.N. V1 on November 14, 2026, using the OpenAI cl100k tokenizer. The 322 vs 19,598 token comparison is from the hosted run. The comparison is the memory sent into the model, not the question or the reply. Prices are official list input rates as of November 13, 2026, with no prompt cache. Shorter conversations save fewer dollars. Your actual bill depends on the model you use and how often you ask. We count whether the part of the chat that held the answer came back, not model-graded answers. This is not a guarantee.
How it works
Add, keep, find.
Add
Bots write short notes as they work. You can also sync Google Drive files you pick: text only, and the originals stay in Drive.
Keep
A rules-based extractor turns new notes into facts. The newest fact wins; replaced facts stay hidden unless a bot asks about the past.
Find
A lookup returns only the facts that matter from your team's own store, so the bot reads only what it needs. No AI model writes or rewrites your memory. The only model is a small search model on your own computer.
Why teams pick it
Honest memory, a flat price.
Memory uses local rules and search. We never call a model to write a note, extract a fact, or fill a gap. That is why the price can stay flat.
Starter and Team are a simple monthly price for the account and the bots that use it. No per-write, per-lookup, per-token, or per-end-user fee. Sign in to subscribe. Billed monthly. Cancel anytime.
We return only what is already in memory. If a fact is missing or only partly there, that is all your bot gets back. Your bot should say so, not fill the gap.
Several bots on one team share one brain. Each bot has its own key, which you can pause or revoke. Other teams cannot see yours.
Use cases
What teams use it for
Several bots share decisions, status and setup facts, so none of them starts from a blank page.
A support bot can remember a customer's preferences and past answers across sessions. Those notes still live in the team's shared memory, so other bots on the same team can read them too. An optional end-user id can narrow a lookup; it does not wall the notes off.
One dashboard, one Drive connection and one invite link for every bot you run.
A bot that serves many people still counts as one bot. There is no fee per end user.
Works with
What connects today, and what is next
- Bots that can follow the invite sentenceMust open a link, store a key, and run a short terminal install. See Setup for examples.Available now
- Google DriveFiles you pick. We store their text only: PDF, Word (.docx), Google Docs, text and markdown.Available
- TypeScript and Python packagesInstall with
npm install @telonyte/brainorpip install telonyte-brain. Package pages: npm and PyPI. They answer from a local copy on your computer and show Hyper speed on or off.Available now - MCP serverComes with the client. It is not an npm package.Available now
- LangChain and LlamaIndex adaptersInstall with
npm install @telonyte/brain-langchain,npm install @telonyte/brain-llamaindex,pip install telonyte-brain-langchain, orpip install telonyte-brain-llamaindex. Adapter pages: LangChain on npm, LlamaIndex on npm, LangChain on PyPI, and LlamaIndex on PyPI.Available now
Measured on our benches
Numbers we have measured
These are lab measurements from our own benches, shown as static figures, not a live counter or status feed.
Notes are stored as written. The only model is a small search model on your own computer, and it never changes a note.
About 322 tokens on the hosted 70-question long-chat test, versus 19,598 for the full conversation. The part that held the answer was in the result 58 times out of 70 with the local search model on, and 50 times with the hosted engine only.
No per-write, per-lookup, per-token, or per-end-user fee. Sign in to subscribe. Billed monthly. Cancel anytime.
Lookup result size is from our own measurement, not an official benchmark result. The 322-token figure is from the hosted 70-question run. On the full LOCOMO set of 776 questions, the part that held the answer was in the result 70.2% of the time with the local search model on (in the top 5 results 70.23% of the time, and in the top 20 87.63%). On LongMemEval, it was in the top 5 results 91.53% of the time, and in the top 20 96.61%. Lookup times here are measured in-process on the server and do not include your network; timed recalls over the internet are in the speed demo. We do not publish a millisecond guarantee as memory grows. These figures measure whether the part of the chat that held the answer came back, not model-graded answers. A test of full answers with a model is still in progress.
Pricing
Starter or Team. No usage fees.
Starter is $11 a month with the first bot included. Team is $19 a month for 10 bots. Extra bots are $2 a month on either plan. Starter is cheaper for 1 to 4 bots; at 5 bots both plans are $19; Team is the better pick from 6 bots up. You choose. We will not switch you automatically. Sign in to subscribe. Billed monthly. Cancel anytime.
See pricing| Bots | Starter | Team |
|---|---|---|
| 1 | $11 | $19 |
| 5 | $19 | $19 |
| 10 | $29 | $19 |
| 100 | $209 | $199 |
Privacy
Plainly stated
Each team's notes live in their own isolated storage. One team cannot read another's. Bots on the same team share that team's memory.
We only read the files you pick and store their text. The original files stay in your Drive.
Pause or revoke one bot's key without touching the others. Revoking a key does not delete the notes it wrote.
We do not train AI models on your note content.
We do not hold SOC 2 or HIPAA certifications today. Formal audits are planned for later. See the draft Privacy Policy.
FAQ
Questions
What counts as a bot, and what is an end user?
A bot is one identity with its own key. The people a bot talks to are end users. A bot that serves many end users still counts as one bot.
How do I install it today?
Sign in to the dashboard, copy the team install sentence and paste it to a bot that can open a link and run a short terminal install. Examples include Grok Bot, Claude Code, and Cursor. Step-by-step walkthroughs are on the Setup page. The JavaScript and Python packages are published too: npm install @telonyte/brain and pip install telonyte-brain. LangChain and LlamaIndex adapters are also published: npm install @telonyte/brain-langchain, npm install @telonyte/brain-llamaindex, pip install telonyte-brain-langchain, and pip install telonyte-brain-llamaindex.
Does each bot only see its own notes?
No. All bots on a team share the team's memory. A key identifies the bot and lets you pause or revoke it. An optional end-user id can narrow a lookup, but it is not a separate access control.
Does B.R.A.I.N. answer questions itself?
No. B.R.A.I.N. is memory your bots use, not a bot to talk to. A lookup returns only what is already in memory. If a fact is missing or only partly there, that is all your bot gets back. Your bot should say so, not guess.
How do I subscribe, and when?
You can connect a bot as soon as you have a team. Sign in, pick a plan, and subscribe when you are ready. Billed monthly. Cancel anytime. If a trial is shown in the dashboard, it will say how long you have.
Where does Drive data live?
We store the text of the files you pick in your team's isolated storage. The original files stay in Google Drive.
Who do I contact?
Use Contact us, or open the help button at the bottom right of any page to ask a question or submit a ticket. That help button is not live chat. We read messages and tickets and reply by email.
AI doesn't know what you want until you give it a B.R.A.I.N.
Sign in to get your team install sentence.