From your own record to a helper that works like you: what we read, what we keep, what we throw away, how it acts, and how we check it against you.
Search, or retrieval: the documents that match the words of a question. Fast and literal. It finds what was written, not what you would do.
A knowledge graph: what relates to what. The buyer, the account, the orders, the terms. It connects the facts. It does not choose between them.
Your engram: your decisions, your rules, your preferences and your way of saying things, each traced to where you said it. The layer that turns facts into what you would do next.
Only your own words. Each source is one you choose, in writing, and can disconnect.
Your replies and the threads they answer. The reply is the source. The thread you answered is context, read only to understand the reply.
What you asked for, what you blocked, and why. The clearest record of an engineer's rules that exists.
Notes, decision logs, the edits you made to drafts. A correction is the most valuable sentence you write.
Your side of meetings and chat, where you connect them. The hardest register to learn, and the one you use most.
Not read: mail you received, notifications, newsletters, anything automated, and anything you did not connect.
Step through one thread. The buyer's name never reaches the engram; the rule in Sam's reply does, citing the reply.
Hi Sam, we'd like to reorder 40 cases for the three Harborlinethe restaurant group locations. Can you do 15% given the volume? Marcus at Lakeside Supplya contact at a distributor mentioned you offer 60-day terms to some accounts.
Context only. Not stored.
Thanks Priyathe buyer. I never go past 10% on reorders, volume earns terms rather than margin. I can do 8% and move Harborlinethe restaurant group to 60-day terms. Shall I confirm the 40 cases?
The source. Its body is not stored either; only the span the node cites.
Never past 10% on reorders. Volume earns terms, not margin.
Source: Sam's reply, 12 March last year · characters 13 to 86
Valid from 12 March last year
Readable by Sam. Deletable by Sam. The only thing kept.
It learns from the sources you connect and from every correction you make. Nothing ambient. Nothing you did not switch on. Step through one correction, then keep or discard what it learned.
A buyer asks for 15% off a reorder.
Happy to offer 15% on this reorder.
15% 8% on 60-day terms
Never past 10% on reorders.
Learned from Sam's edit on 3 October
Never past 10% on reorders.
Two weeks later, another buyer asks for 12% off a reorder.
I can do 8% on 60-day terms, which is what I offer on reorders.
Cites Sam's edit of 3 October.
No precedent for this. Asking Sam.
Nothing to cite: Sam discarded the node.
Drafted by Sam's AI. Sam approves before it sends.
Between sessions it consolidates what it learned, and you review the list each week: what it proposes to keep, with the source of each.
Replies, summaries and notes in your voice, each labelled as drafted by your AI, waiting for you.
The situation in one line, the recommended action, one or two alternatives, the cited rule or precedent, and Approve, Edit or Decline. Answered in seconds.
Nothing goes out, gets paid or gets promised without your OK, unless you set a rule that allows it, inside limits you set.
Every message it sends says it was drafted by your AI, and carries a machine-readable marker. Disclosure is built in, not optional.
We hold back decisions you already made, replay each one, and count agreement, reported as a fraction of your own consistency, with its margin. The method is on its own page; the number comes when the test has run.
| Dimension | Memory built into assistants | Clone platforms | Your engram, with Duplikate |
|---|---|---|---|
| What it learns from | Your chats with that assistant | Documents and recordings you upload | Your own record: sent mail, reviews, notes, corrections, and only the sources you choose |
| What it keeps | Facts about you, as it summarises them | Your content, for retrieval | Decisions, rules, preferences and voice, each citing where you said it |
| Who else is in it | Whoever you mentioned, as said | Whoever is in your uploads | Nobody by name: people and companies become role types at the door |
| Is it measured against you? | No published score | No published score | Yes: held-out decisions, replayed, reported as a fraction of your own consistency |
| What it does | Answers you | Answers your audience | Drafts, triages and prepares decisions you approve |
| Leaving | Export, in some | Varies | Export everything with its sources, any time; 90 days' notice if we ever close |
Based on the public pages of products in each group, read on 24 October 2026. A sourced, product-by-product comparison follows in Learn.
You want an assistant that acts with no review. Ours asks first, by design, until it has earned a standing rule inside a limit you set.
You want a video or voice double. Duplikate writes. It does not imitate your voice or your face.
You won't share your own writing. There is nothing to build from without it, and we never build from someone else's.
You'll talk to a senior engineer, not a sales rep, and leave knowing what your engram would be built from and how it would be measured.