Option 1
Paste it into a public chatbot
Client data leaves the building.
HTHolt Teams · Design partners wanted
A private AI platform that runs inside your company and only shows each person what they're allowed to see. Nothing reaches an outside model unless you allow it.
In design now. Building it with 3 to 5 companies.
01Why
Option 1
Client data leaves the building.
Option 2
Safe, but the work stays manual.
Holt Teams
Useful AI, on your servers, with IT in control.
02How
Personal data is processed on your side. Only masked text goes to the frontier model your admin allows.
03Orchestration layer · included by default
Plug in Claude, OpenAI, Gemini, Mistral or any model with an API. Holt decides what goes where, by your rules.
Every request goes through the orchestrator. Nothing to install.
Personal data is handled on your side. Reasoning uses the model you choose.
Names and numbers swapped out before a cloud call, put back on return.
Change models without moving your data or retraining anyone.
04Security
A box in your office, or your own cloud account for bigger teams. Hosted by us if you have no IT.
Can't open it in the source system? Holt can't show it.
Each person's memory lives on their own laptop.
Nothing opens up without admin approval.
Encrypted before they leave your server.
Updates are signed and opt-in. On your box or your cloud, I can't reach it.
05Memory
06Examples
Yours will be different. That's what design partners are for.
Digging through years of candidate notes by hand
Ask once, get matching candidates with sources, then a draft client report
Customers go quiet and nobody notices in time
Lapsed customers flagged from orders and reviews, win-back message drafted
More workflows it is designed for
Client names and account numbers can never go into a public chatbot, so the questions stay manual
Names and numbers are masked on your side, the model you allow does the reasoning, the answer comes back with sources
Policy questions land on whoever is nearest, and salary files are one wrong share away
Everyone asks the same assistant, and it only answers from what that person can already open
What one person learns about a client stays in their notes and leaves when they do
Useful notes are suggested for sharing, the person can veto, an approver signs off, and the team keeps it
Patient records cannot leave the building, so the AI tools everyone else uses are off limits
It runs on your own hardware, reads records where they live, and nothing reaches an outside model unless you allow it
07Status
The open-source core, Holt, is live and past 5,000 downloads. Holt Teams is next.
08Benchmarks
Holt Teams is being built on the same memory engine as Holt. Here is how that engine scores on the public memory benchmarks.
500 questions about a user’s past chats. Graded by GPT-4o, the benchmark’s own grader. 89.7% on the 399 questions never used for tuning.
1,540 questions across long two-person conversations. Graded with the original, stricter LoCoMo prompt.
700 questions over conversations of about a million tokens each, too long for any context window.
About 6,000 tokens of memory per question. Search and storage run on your own machine with a local model. Answers came from Claude Sonnet 5.5. LongMemEval was graded by GPT-4o, LoCoMo and BEAM by Claude Opus 5.5.
All tuning used the same 101 LongMemEval questions. LoCoMo, BEAM and the other 399 were never used to choose anything. Other systems’ numbers are their own published results, with different models and graders. *Mem0’s LoCoMo grader is more lenient than the one used here.
See every answer and rerun it yourself →09Deal
What you get
What I ask
10Apply
Seven fields, two minutes. I read every one and reply personally.
Company size