Tutorial 26 · Memory in depth: training JMAX on what's yours

What the video says

1. Memory is what turns JMAX into your assistant rather than just any assistant: this is where it keeps what you talk about, and where you load your documents. It lives in the menu, under Management and administration.
2. At the top, persistent memory. Everything you say, every step JMAX takes and every spoken sentence is stored with its meaning, and none of it is lost when you close the application.
3. The status is the health check: where it is stored, how many episodes it holds and how many it wrote this session. Watch Pending and failed above all: if they are not at zero, something is not being saved.
4. Below, training with documents. It takes PDF, Word, Excel, Markdown and plain text, and also photos and videos, which it turns into memories with a date and a description. Twenty-three formats in all.
5. Before uploading anything, the workspace. This is the decision that changes the result the most: group by subject, not by origin. Contracts, Manuals, Minutes. One workspace per topic, with names you would say out loud.
6. Pick the files and press Upload and train. A tip: start with few and good ones. Twenty documents you actually use are worth more than five hundred nobody reads, and JMAX answers better when the workspace is clean.
7. Each workspace shows how many documents it holds and how many fragments they were split into. The fragment is the unit JMAX searches, so a long document contributes many: that is why each workspace should have one clear topic.
8. And here is the important exception. If a workspace carries people's data, customers, employees, patients, tick the Customer data box. The names are replaced by pseudonyms before any model ever sees them.
9. Tick it before uploading the documents, not after: the replacement happens while training. The document button downloads the evidence that it was done, which is what an audit will ask you for.
10. The personal data firewall watches every turn: when it spots sensitive data, it handles that turn on the local brain instead of sending it out. The counter tells you how many times it has done so.
11. Encrypting the databases protects the whole computer: if someone walks off with it, they cannot open your conversations or your documents. The key stays on this machine and nobody has to remember it.
12. The right to be forgotten. Type a person's name, email or ID and press See what's there. Look first, decide after: permanent deletion is a second button, and there is no undo.
13. How long each thing is kept, in days. Zero means no limit. Out of the box, memory for two years, ratings for one, agent work for half a year and teaching captures for a month.
14. Save periods sets the rule going forward. Apply now does the cleanup in one go, on what is already stored. Set this on day one, with your data policy in front of you.
15. The vector database is where the meanings live. Local needs nothing installed and covers almost everything. Move to an external server only if there are a great many documents, or several machines share the same memory.
16. Search the history searches by meaning, not by date or exact word. Ask the way you would speak: what did we agree with Acme about the payment term. It is for recovering what you know was said and can't place in time.
17. And View 3D map opens the knowledge graph. This is the part worth showing when somebody asks what exactly JMAX knows.
18. This is everything it knows, at a glance. In blue the workspaces, in gold the documents, and around them the ideas each one was split into. The red nodes are the topics that came out of conversation, not out of the files.
19. The yellow lines are the important part: they join what you talked about with what your documents say, by similarity of meaning. You can see the conversation about payments touching the master agreement and the service agreements.
20. The search box narrows the map to what matters. Type payment and you are left with the topic, the contract term and what was agreed. It is the quick way to answer where an answer came from.
21. And clicking a document opens its record: the plain-language explanation, the workspace, the original file, who uploaded it and when, and the exact content fragment by fragment. That is what you show when a client doubts an answer.
22. To make this pay off: one workspace per subject, few documents and good ones, tick the ones carrying people's data before uploading them, and set the retention periods on day one. Then JMAX answers with what's yours, and says where it came from.