---
title: Conversation and topics
summary: A person has one continuous conversation with the assistant in each business. The assistant divides it into topics, keeps earlier topics in short, and can search what was said before.
tools: []
related: [staff-assistant, assistant-mode, confirmation-cards, feedback-to-team]
search: [chat history, memory of the assistant]
---

# Conversation and topics

## One conversation

In each business a person has a single conversation with the assistant. It does not
end and is not started over. Only that person sees it. A key of the business or a
program cannot read or write it.

The conversation goes back in pages: **Earlier messages** shows older ones.
Each message of the person is one run, so while an answer is being written the
next message waits (see [Run](../runs/run.md)).

## Topics

When the subject changes, the assistant starts a new topic. It names the topic in
the language of the person, and the name shows as a line between the messages. In
the app the latest topic's name is also the title of the screen. The assistant makes
the topics itself.

## What the assistant keeps of the past

The assistant has no memory outside the conversation, and it builds each answer
from the conversation itself.

- The current topic comes as it was written, as far as it fits.
- Earlier topics come as short summaries, a few of the latest. The summary of a
  topic is made when the next message arrives. An older part of a
  very long topic is shortened the same way.
- Something older still the assistant finds by searching the conversation. The step
  shows as "Searched our conversation".
- What a tool returned in an earlier run is not kept. If the assistant needs it, it
  reads the record again.

What should stay for clients is written in the knowledge (see
[Knowledge entry](../knowledge/knowledge-entry.md)).

## Suggested next questions

After some answers the assistant offers one or two short questions that the person is
likely to ask next. They appear above the input line, and a tap sends one in the
person's words.

## Under an answer

- **Copy** copies the text as you see it, without the bold and list marks.
- **Answer again** runs the same words once more, on the last answer, if the
  assistant changed nothing in it.
- **What I did** and **Where it comes from** show the steps and the sources. A step is
  named for what it does: a read as "Looked at the contacts", a change in the words of
  the decision button, like "Change the online admin’s script". Next to a step there is
  a note if it was refused ("Didn’t work"), only shown beforehand ("Shown beforehand"),
  or it is not known whether it worked.
- A line "Change the online admin’s script · Done" stands under the answer for every change
  that went through. Talemo puts it there, not the assistant: the assistant can be wrong
  in words, the line cannot. No line means nothing was changed, whatever the text of the
  answer says.

## How an answer looks

Bold, italic, bulleted and numbered lists and links in an answer are drawn, not shown as
signs. A heading looks like a bold line. A link opens only if it is an `http` or `https`
address; it opens outside Talemo: in a new tab on the site, in the browser in the app.
There are no tables: a table written with marks stays plain text.

## Chats that were ignored

What the assistant has already told the person about a client's chat stays in their
conversation when that chat is ignored. The chat's own messages and the records of
runs that carried its words are erased.
