Your agents remember the whole conversation, however long it gets. When a conversation grows past what an AI model can read in one go, HALO condenses the older part into a summary and keeps the most recent messages word for word. The agent reads the summary plus the recent messages, so it still knows what was discussed, agreed, or asked for earlier.
You do not configure this. It works the same way for every agent, in chat and in voice conversations.
How it works
-
HALO keeps the full transcript of every conversation. Nothing is deleted. The Conversations page always shows every message.
-
Each time the agent answers, HALO checks how much history the model would have to read. Short and medium conversations are sent in full.
-
When the history grows past the limit, HALO asks an AI model to write a summary of the older messages. The summary covers the customer's requests, the answers given, the data collected, and anything still open.
-
From then on, the agent reads the summary followed by the most recent messages. As the conversation continues, HALO folds the newer messages into the summary, so there is always one summary and a short tail of recent messages.
-
The summary is internal. Your customer never sees it.
What to expect
The summary keeps the facts that matter for the rest of the conversation: what the customer wants, names, numbers, choices made, and what the agent promised. Exact wording from early in a very long conversation can be lost. If a tool must use a value later in the conversation, such as an order number or an email address, store it in a context value when you collect it instead of relying on the agent to recall it from the history.
If the summary cannot be written, for example because the AI model is temporarily unavailable, the agent keeps using the full history and HALO tries again later. The customer notices nothing.
Reference
|
Setting |
Value |
|---|---|
|
History sent in full up to |
about 16,000 tokens or 512 messages |
|
Kept word for word after summarising |
the most recent 16 messages, about 2,000 tokens |
|
Summary length |
up to 2,000 words |
A token is roughly three quarters of a word. These values are the same for every account and cannot be changed in HALO Studio. The AI model used for summaries is listed on AI Models in HALO.