Skip to content

Chat History

The agent stores all chat conversations in Splunk's KV store, a key-value database that provides persistent storage for application data. Chat history is maintained in the chat_history collection on the Search Head where the app is installed, and is replicated across all members in Search Head Cluster deployments.

Storage and Access

Chat history is stored per user, ensuring privacy and security. Each user can only access and manage their own conversations.

Users can:

  • View their complete chat history, and search it by title
  • Resume previous conversations
  • Rename and delete individual conversations
  • Export a conversation to PDF, or to a JSON file that can be imported again later

Conversations are titled automatically from their first messages, and can be renamed at any time.

A chat also remembers the model it was created with, and selects it again when the chat is reopened.

Context Window

Each conversation has a limited context window, shown as a fullness indicator at the bottom of the chat sidebar. Hover it to see the tokens used against the selected model's context window.

Once a conversation approaches its limit, the indicator turns red and it's best to start a new chat or compact the current one.

Compacting a Conversation

Compacting summarizes a long conversation so work can continue without its full history.

Select Compact Context from the 3-dot menu at the bottom of the chat sidebar. The agent writes a summary of the conversation so far and opens it as a new chat named after the original, with (compacted) appended. The original conversation is left untouched.

The dashboard you're currently working on is carried over in full, so you can keep refining it in the compacted chat without regenerating it.

Tip

Compacting carries over the important details, but not everything. If a specific search matters for what comes next, mention it again after compacting.

Infinite Chat

A conversation that outgrows the context window doesn't stop working. With Infinite Chat, the agent slices messages out of the middle of an over-long conversation so the request still goes through, keeping the opening question and everything from your latest message onward. The saved chat itself is never modified - the messages are only left out of that one request.

The middle is sliced because that's where a model is least likely to miss what it lost: language models attend most strongly to the beginning and the end of a long input, and least to what sits in between (Liu et al., Lost in the Middle). The framing of the task and the turns you're actively working on are what stay.

A warning icon next to the fullness indicator shows how many messages were sliced out.

Note

While Infinite Chat keeps you moving, it doesn't give the agent more room. Once it kicks in, the agent no longer sees everything that was said, so compacting the conversation or starting a new chat is still the better move.

KV store Requirements

Warning

Chat history functionality requires KV store to be enabled on the Search Head.

If KV store is disabled or unavailable, the agent will not be able to store or retrieve chat history. In this case, users will not see their previous conversations, and new conversations will not be persisted between sessions.