Work when capacity is available
When the profile has Decision and Dream text resources, topic classification can start as a message arrives, alongside its processing. Chat does not wait for it. Human corrections retain priority; classification grants no additional access and does not repeat a search that has already run.
Dream uses available periods for background maintenance. It runs within configured duration and retry limits, yields to priority activity such as voice and stores receipts for its work.
Consolidate useful context
Classification, memory extraction and process projections help connect completed work to durable knowledge. Maintenance findings identify duplicates, contradictions or ageing. You can inspect decisions instead of assuming that background activity improved every result.
Learn from evidence
Procedural learning considers task outcomes, tool evidence and human corrections. A candidate retains its score and supporting tasks before activation. One successful attempt is not enough to establish a reliable procedure.
Keep the scope clear
Dream is background processing, not model retraining or unrestricted self-modification. It operates on authorised information and configured policies. The learned-skill feature is disabled by default and should be evaluated on the work it is meant to improve.
Recorded decisions, one mechanism at a time
The fourteen registered mechanisms classify and extract knowledge, project processes, maintain memory, repair document links, index files and prepare thumbnails. Four attachment analyses remain optional, and procedural learning follows its selected mode. Maintenance is sequential; early message classification has its own parallel trigger.
Extraction proposes CREATE, LINK or IGNORE. Adding a source does not rewrite the memory content. Receipts retain prepared output and its application so recovery can reuse an existing model result.
Keep dates that matter for recall
When a source describes an appointment, birthday or recurring habit, extraction can propose a partial temporal anchor with its timezone. A relative date refers to the source message, not the time Dream processes it. If the date is ambiguous, the anchor remains empty; a date mentioned in a report does not automatically become a reminder.
These conditions follow memory acquisition and revisions. They support contextual recall without creating a notification or calendar event.
Observe before learning
Modes are off, observe and learn. Observation retains a proposal; learning can apply it. A score and evidence from multiple distinct tasks—three by default—must support activating a procedure. Positive evidence, corrections and failures remain inspectable.
The skill library distinguishes these procedures from ordinary SKILL.md packages.
Describe attachments when enabled
Four independent options, off by default, fill empty descriptions: text and convertible documents, documents requiring a document model, images, and video audio tracks. Text PDFs and supported office formats can be extracted and summarised; scanned PDFs require an appropriate document resource.
Dream handles one attachment per mechanism pass, retains receipts and preserves existing descriptions. The video option transcribes sound; it does not analyse frames. Unsupported files may remain without a description. Repairing document links is separate from these generative analyses.
From monitoring to file indexing
Tracking combines indicators and activity; History lets you inspect operations and receipts; Indexing shows file scans for the selected agent. Launch an eligible scan, inspect its partial coverage, cancel it or retry repairs. Discovery depends on the provider and connection settings.
The catalogue and thumbnails are prepared without AI calls. Optional analyses can enrich descriptions while preserving personal notes. Available actions can also be launched from an authorised memory detail view. Configure file indexing.



