Skip to main content

Agents and conversation history

Conversation history is stored server-side and can be reviewed or deleted; in the open-source build an agent is an internal binding, not something you create.

Two separate things share this page, and they are worth telling apart up front. Conversation history is a real, everyday feature. "Agent" is an internal binding in the open-source build, not something you create — the details are in the second half so you do not go looking for a page that is not there.

Where conversations live​

Every turn is written to the database as it happens, so a conversation survives a page reload, a different browser and a restart of n9e. Two tables hold it:

TableWhat it holds
ai_assistant_chatOne row per conversation: its id, its owner, its title, when it was last touched
ai_assistant_messageOne row per turn: your question, the answer, and the tool calls behind it

Both bodies are compressed, so the rows are smaller than the transcripts suggest.

Conversations are strictly per user. Every read and write checks that the requesting account owns the conversation, with no administrator override, so one person's chats are not visible to anyone else — including through the API.

An /a2a conversation is the same object: the contextId a caller passes back is the chat_id, so an agent talking to Nightingale through A2A appears in that user's list alongside what they typed in the browser.

What you can do with one​

The Conversations section of the Nightingale AI sidebar lists yours, newest first, grouped into Today / Yesterday / older, with a box that filters on title. Hovering a row reveals a menu:

  • Rename — inline. Titles are generated from the first question otherwise, and renaming pins yours so the automatic title never overwrites it;
  • Share — copies a link that opens the conversation read-only. The server still checks ownership on every read, so the link works for you and nobody else; treat it as a bookmark to a read-only view, not as a way to hand a transcript to a colleague. To do that, copy the text.

Deleting one​

There is no delete button. The endpoint exists; nothing in the interface calls it. Until that changes, deleting is an API call with your own token:

curl -s --request DELETE \
  -H 'X-User-Token: YOUR_TOKEN' \
  http://127.0.0.1:17000/api/n9e/assistant/chat/YOUR_CHAT_ID

The chat id is the last path segment of the conversation's URL. Deleting removes its messages as well, and — the ownership check again — only works on your own conversations.

There is no retention policy and no TTL. Nothing expires conversations, so on a busy instance these two tables only grow. If that matters, script the call above; they are also ordinary tables that your backup and archival routine can treat like any other.

What an "agent" is here​

There is an ai_agent table, and rows in it bind a use case to an LLM config and a list of skills. In the open-source build that is much less than it sounds:

  • The only use case is chat, and exactly one enabled row for it is ever consulted;
  • A row is created for you at startup — default-chat-agent, bound to no specific model — so the assistant works out of the box;
  • There is no system prompt and no tool selection on the row. Both are compiled into the product; nothing on an agent changes how the assistant thinks or what it may call;
  • The skills bound to the chat agent are ignored on the ordinary conversation path. Skills are discovered from the catalogue and loaded on demand instead — see Install, manage and write Skills;
  • There is no menu entry, no permission point of its own, and the list page that does exist has had its create button removed. It is administrator-only and reachable by typing the address.

So: do not build a workflow on named agents in the open-source build. If a page or a screenshot elsewhere shows agent management as a feature, it is not this edition.

What to change instead of an agent​

What you actually wantWhere to do it
A different model for the assistantMark a different LLM config Default — it wins over any agent binding anyway (Configure an LLM provider)
Different behaviour for a class of questionWrite a skill; its description decides when it applies (Install, manage and write Skills)
More tools for one procedurebuiltin_tools in that skill's frontmatter
A longer diagnostic chainmax_iterations in that skill's frontmatter
Restrict what the assistant can reachThe asking account's role and business groups (Permission inheritance and RBAC)

Next​

On this page