> ## Documentation Index
> Fetch the complete documentation index at: https://docs.usealmanac.com/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> This handbook preview targets personal CLI 0.11.1. Check installed command help and the matching backend schema before using examples with another release. Examples with named people and placeholder IDs are fictional.

# Work between conversations

> Shadow and Dream are Almanac working between conversations.

Shadow and Dream are Almanac working between conversations. They use the same private knowledge, tasks, projects and sessions as the conversational agent. Their assignments differ in scope: Shadow reconciles changed conversations and current work, while Dream builds a broader understanding across the person's sources and history.

An email asking for a reply is a starting point for investigation. A later message may show that the person already replied, that a teammate fulfilled the request, or that the outcome changed. Background work should establish what remains useful now before adding a task or interrupting the person. [Tasks and continuing work](/almanac/tasks) explains where that remaining outcome belongs.

## Shadow: reconciling change

Shadow examines pending conversation changes, follows relevant originals and reconciles saved understanding and work. Some changes need an edit or an authorized preparation step; others need no change. Reviewing an item and deciding that nothing should be done is still real review.

The review interface returns a metadata revision and a content revision. Shadow reads and acknowledges the versions actually examined, using the version-bound review operations. A familiar conversation can still have pending metadata or content changes. A final result page still needs review even when it says no further page exists. The operational [review guide](/work/review) explains these checkpoints.

Those receipts have a narrow meaning: they record review of particular versions. They do not establish that a task is complete, that every provider notification has been handled or that the user saw a message. Shadow must verify any consequential changes separately.

## Dream: connecting the wider history

Dream's current assignment calls for thorough research across connected accounts, conversations and existing knowledge. It maintains a working research notebook containing discovered subjects, source references, uncertainties, further investigation and saved outputs. Listing an account does not count as inspecting its contents, and an inaccessible source remains an explicit coverage gap.

This assignment is broader than answering a question in chat. Dream accounts for the user and every discovered person and organization with a page, and creates project overviews for identified projects. It deduplicates identities and develops those accounts with detail proportional to evidence. A concise subject can stay concise; an uncertain identity must not acquire invented details. General advice about selective page creation does not excuse silently dropping subjects from Dream's research inventory.

The notebook connects discovery to saved understanding. It may begin with an uncertain relationship and end with a cited person article, a project connection or a justified conclusion that no task remains. [Preferences](/almanac/preferences) are one form of that understanding: reasons, exceptions and changing circumstances matter more than turning observations into universal rules.

## Protecting the person's attention

Background activity does not automatically warrant a message. Shadow is instructed to investigate and prepare first, then interrupt for useful ready work, a timely decision or a material consequence the person would otherwise miss. Repeating an alert or offering to investigate it is usually insufficient. Later evidence may resolve an earlier failure before the person needs to hear about it.

A quiet response does not finish pending reviews. A delivered update does not prove the underlying work is complete. Background agents should leave useful saved state, preserve unresolved gaps and verify the effects they report. Their external actions follow the same [permission boundaries](/almanac/initiative-and-permission) as work in chat; background execution does not add authority.