Context and History

“The conversation is stored on this device” and “the model can currently see the entire history” are different statements. A conversation may be retained for a long time, but every model request has a finite context window.

What the current context contains

A model request typically combines:

User messages and assistant responses that remain active in the current conversation.
Safe results from completed tool calls.
Necessary information about the current project, permissions, and run state.
Instructions from selected or activated skills.
Historical summaries and exact-retrieval references produced by automatic compaction.

The interface timeline is a display and is not identical to the model context. Hidden reasoning is not presented as ordinary chat content. To preserve continuity with the same provider, the runtime may still retain necessary provider-specific continuation state. Credentials, private storage paths, and data that should not be sent do not become ordinary context merely because they appear in an internal process.

View context usage

If the current build includes the context indicator, enable “Show context-window usage” under “Settings → General → Composer.” The composer will display the percentage and number of tokens used.

This value is based on the context assembled by Captain Who and the configured model window. If you enter a window larger than the provider actually supports, the API may still reject the request.

What automatic compaction does

When history approaches the safe capacity, Captain Who:

Selects a complete prefix of older history.
Asks the model to produce a structured summary.
Saves the summary and a small set of exact references.
Replaces the old prefix with the summary, then continues the task.

The timeline may display “Automatically compacting context,” “Context automatically compacted,” or a failure state. Compaction aims to preserve the task's meaning, decisions, unresolved questions, and important evidence. It does not guarantee verbatim retention of every older detail.

For a very long conversation, ask the agent to write “current state, key decisions, remaining work, and verification results” to a project file, then start a new task. This is usually more reliable than relying on unlimited history.

Retrieve history

Captain Who stores bounded audit history for completed tool activity. Some large text captures can also be retrieved precisely page by page. When needed, the agent can use conversation-history capabilities to find earlier results; the interface may display “Recalling.”

The boundaries include:

Each capture and upstream tool has size limits.
Binary data, secrets, and hidden reasoning are not included in ordinary history archives.
History cannot recover content that an upstream source already truncated.
Local conversation search uses keywords; it is not vector-based semantic memory.

Do not treat conversation history as an unlimited, lossless, cross-device knowledge base.

Context when switching models

When the new model uses a different API provider or compatibility rules, Captain Who requires “Compact and switch.” This converts provider-specific state from the previous service into a summary the new model can safely consume, rather than sending incompatible internal fragments directly.

You cannot switch while a task is running or waiting for approval. If conversion fails, the original model selection and history remain unchanged. Resolve the error before trying to switch repeatedly.

Branch, archive, and delete

Continue in a new task: Create a history snapshot from a completed response or model-conversion boundary. The old and new conversations do not remain synchronized.
Archive: Remove a conversation from the main list while keeping it available for restoration in Settings.
Permanently delete: Delete local conversation content and associated authorizations. This cannot be undone from the interface.
Remove project: Delete the project's conversations and attachment records from Captain Who without modifying files in the project directory.

Conversation attachments and managed artifacts have separate authorization and retention mechanisms. There is currently no unified permanent artifact library or user-configurable long-term retention commitment. Explicitly save important files to a directory you manage.

Data safety

Conversations and settings are stored primarily on your device, but context sent to a model reaches the API you configure. Do not place secrets in a conversation unless the model needs to process them. There is currently no public cross-device synchronization or one-click export of all history.