What's new for you
Broad quality release centered on plugin loading boundaries, memory/dreaming reliability, and local-model ergonomics: scoped plugin activation by manifest intent, bundled LM Studio provider with onboarding and embeddings, Active Memory plugin and Codex/exec-policy/gateway command discovery refinements, and many fixes across startup sequencing, chat command catalog, dreaming promotion/diary behavior, keepalive, queueing, and channel delivery.
- Plugins — activation/loading narrowed to manifest-declared needs with stronger trust boundaries
- Local models — bundled LM Studio provider with runtime discovery and embeddings
- Memory — Active Memory plugin plus dreaming and memory-wiki reliability improvements
- Gateway/UI — startup gating, slash-command catalog in composer/palette, update flow resilience
- Channels — WhatsApp, Telegram, Discord, Matrix reliability and delivery fixes
- Security — shell/interpreter hardening, approval checks, and safer defaults
Release highlights
| Feature | Details |
|---|---|
| Plugin loading boundaries | CLI, provider, and channel activation now follows manifest-declared needs so startup and runtime avoid loading unrelated plugin runtime, preserving explicit scope/trust boundaries. |
| Bundled LM Studio provider | Adds LM Studio as a bundled provider with onboarding, runtime model discovery, stream preload support, and OpenAI-compatible memory-search embedding routing for local/self-hosted models. |
| Active Memory plugin | Optional memory sub-agent before the main reply with configurable context modes, /verbose inspection, and transcript persistence options for deeper tuning and debugging. |
| Gateway startup and command UX | Defers scheduled services until sidecars finish, gates history/model listing during startup resume, and loads live slash-command catalog into chat composer and command palette. |
| Dreaming reliability pass | Fixes managed-heartbeat dedupe, promotion thresholds, timezone clarity in diary entries, transient narrative cleanup, and self-ingestion loops in dreaming pipelines. |
| Keepalive and queueing stability | Stops droppable keepalive tick disconnects under backpressure and preserves orphaned active-turn user text so follow-up messages are not silently lost. |
| Channel delivery fixes | Includes WhatsApp media/reaction routing fixes, Telegram approval-lane deadlock prevention, Matrix mention compatibility, and Discord heartbeat reconnect stability. |
Feature deep dives
Learn more about the changes that matter most. Bug fixes and minor improvements are listed on GitHub.
Plugin loading boundaries
Manifest-driven activation narrows startup/runtime plugin loading and preserves trust boundaries.
Read more →LM Studio bundled provider
New bundled local-model provider with onboarding, discovery, and embedding support.
Read more →Active Memory plugin
Optional memory sub-agent before replies with mode controls and debugging telemetry.
Read more →Gateway startup and command catalog UX
Startup sequencing and live slash-command catalog improvements for the Control UI.
Read more →Dreaming reliability pass
Promotion, diary, cleanup, and dedupe fixes improve dreaming consistency.
Read more →Keepalive and queueing stability
Prevents tick-timeout disconnects and preserves follow-up user text during active turns.
Read more →Channel delivery fixes
Delivery-path fixes across WhatsApp, Telegram, Matrix, and Discord.
Read more →Full release notes
The complete changelog with all changes, fixes, and technical details is on the official GitHub release page.
Open v2026.4.12 on GitHub →