Arbor 0.21.0

Arbor 0.21.0

This release repairs long-running agent work and adds first-class local memory while relaxing arbitrary generated-widget UI limits.

Long-running work and Continue

• Streaming model requests no longer fail merely because a reasoning model emits no SSE bytes for two or three minutes.
• Manual Continue waits for the previous WorkManager instance to finish cancelling before marking the message as streaming, eliminating the cancellation race which made Continue appear to do nothing.
• Worker replacement cancellation no longer overwrites a newly resumed message with INTERRUPTED.
• Automatic output continuation rises from three to twelve segments.
• Normal tool workflows can use up to 64 tool rounds and Deep Research up to 128, replacing the previous 8/24-round ceiling which prematurely finalized widget-building sessions.

Memory

• Adds encrypted cross-chat memories stored in the SQLCipher Room database.
• Adds native memory_save, memory_list, and memory_forget tools with visible Working events.
• Injects enabled memories as bounded user-owned reference data, never as instructions.
• Adds Memory settings with global enable, automatic-memory policy, manual add, per-item enable/disable, and delete controls.
• Includes memories and memory policy in app-settings backups.
• Automatic memory excludes transient details and requires explicit consent for sensitive information.

Generated widgets

• Raises arbitrary schema ceilings while retaining bounded resource and security limits.
• Accepts input-style nodes on Home widgets as state readouts/actions instead of rejecting the entire program.
• Removes the six-row renderer truncation.
• Adds bounded multiline text wrapping and ellipsis instead of flattening every line into one clipped row.
• Exposes input actions in the Home-widget action strip.

Verification

Third-party license verification, offline license generation, release unit tests, release lint, and the optimized release APK build passed before publication.