Voice and Memory extend BLXAde with spoken control and durable, scoped memory. Both are configured under Settings and optional onboarding panels.

Voice overview

Voice is a native subsystem. Audio capture and model runtimes stay outside the React webview. PCM is not persisted as raw user recordings in Settings.

Control modes

Typical product modes (names in UI may be localized):

Mode	Role
Local speech recognition	On-device STT (Parakeet catalog models)
Cloud Realtime	OpenAI and/or xAI realtime conversation
BLXFusion	Combined local + cloud arbitration with single capture (local-first arbitration; wakephrase can gate cloud audio)

Exact availability depends on installed models, devices, and provider keys.

On-device extras

Feature	Behavior
Groq Agent Brain	Optional fallback when local Voice Command misses; isolated key; `tool_choice: none` completions
Supertonic TTS	Local EN/DE speech; load/unload without deleting model files
xAI / catalog TTS	Cloud talkback for final on-device answers when configured (typed/off/Realtime paths stay silent as designed)
Dictate clipboard recovery	Off-by-default policy: bounded write-only recovery if Dictate fails (privacy disclosure in Settings)
Wakeword	Optional; details collapsed when disabled
Activation grouping	On-device vs Realtime ownership; push/toggle/app-wide and optional OS-global shortcuts

When enabled, workspace-scoped resume capsules can reuse an xAI provider conversation inside a short inactivity window after project/app reopen. Completed tool outputs stay exact-once across process restart; uncertain effects are never auto-rerun. Revocation clears local state first. Transcripts and raw arguments are not durable free-text stores.

Settings → Voice

Sections include:

Control mode — which voice path is active
Realtime — OpenAI / xAI configuration when keys exist
Activation — how voice starts (shortcuts, wakeword)
Local speech recognition — model download/load state
Microphone — device select, refresh, test meters
Wakeword — optional; details collapse when disabled
TTS + Talkback — spoken replies and policy
Miscellaneous and Privacy
Diagnostics — content-free health (no secret dumps)

Preferences autosave. Model install, device tests, and diagnostics remain explicit actions.

Provider keys

Configure under Settings → Connections:

OpenAI (Realtime / related)
xAI (Realtime when enabled)
Groq (where used as fallback/brain path)
OpenRouter remains the Global Agent path

Keys use the native provider vault (OS keyring), same family as OpenRouter.

Shortcuts

Default voice bindings (rebindable where Settings allows):

Intent	Default
Converse	Ctrl+Shift+R
Dictate	Ctrl+Shift+D
Voice command	Ctrl+Shift+Space
Cancel	Esc (system, read-only)

See Shortcuts.

Permissions

macOS: microphone purpose strings are bundled
Windows: WASAPI privacy prompts
Linux: ALSA / PipeWire host selection

If the mic test fails, fix OS permission first, then retest in Voice Settings.

Memory (CrossMemory)

CrossMemory is the durable memory store for agent-relevant notes with scopes, revisions, and recovery tools — not a second chat transcript UI.

Settings → Memory

Operational console for:

Runtime health and embedding lifecycle
Search (active / tombstone)
Provenance and revision history
Restore, promotion, export
Confirmed permanent deletion

Preferences autosave; destructive and index operations stay explicit.

Danger Zone wipe

Settings → General can clear all CrossMemory scopes after typing the exact phrase:

REMOVE MEMORY

This is separate from clearing recent projects or replaying onboarding.

Notifications and Talkback

The Notification Center keeps privacy-safe durable events (no raw prompts, paths, or credentials). Voice answers may toast/speak transiently without persisting full spoken text into the event store. Configure related options under Settings → Notifications and Voice talkback policy.

Tips

Install local STT models while online before relying on offline recognition.
Keep Realtime keys only if you use cloud voice — local mode does not need them.
Prefer agent tools + Memory for durable facts; do not paste secrets into memory entries.

Voice & Memory