v2026.4.12 — Active Memory + Codex Integration
Released: April 13, 2026 Theme: Memory sub-agent + model integration expansion Breaking Changes: None
Core Highlights
1. Active Memory Sub-Agent
New Active Memory plugin: dispatches a dedicated memory sub-agent right before the main reply, automatically pulling user preferences, ongoing context, and past details into the answer.
Previously users had to explicitly say “remember this” or “search memory” to trigger memory lookup — now it happens automatically. Includes:
- Multiple context modes (single message / recent / full conversation)
/verboseto inspect what memory was loaded in real time- Advanced prompt and thinking overrides for tuning
- Optional transcript persistence for debugging
2. Codex Provider Integration
Codex provider is now fully integrated: codex/gpt-* models route through Codex-managed auth, native threads, model discovery, and compaction; openai/gpt-* stays on the regular OpenAI provider path.
Two parallel paths, no interference.
3. macOS Talk Mode Local MLX Voice
Talk Mode adds an experimental local MLX speech provider: explicit provider selection, local utterance playback, interruption handling, with system-voice fallback when needed.
No audio leaves the machine.
4. LM Studio Local Models
New bundled LM Studio provider: onboarding, runtime model discovery, stream preload, and memory-search embeddings — all supported for local OpenAI-compatible models.
5. Plugin Loading Optimization
CLI, provider, and channel activation now narrows to manifest-declared needs. Startup and command discovery no longer load unrelated plugin runtime. Faster boot.
6. Dreaming and Memory Reliability Fixes
- Dreaming stops re-ingesting its own narrative transcripts
- Short-term memory dedupe uses ingestion day instead of source file day
- Wiki slugs preserve Unicode letters and combining marks — non-ASCII titles no longer collapse
- Memory Palace’s Imported Insights stops showing misleading unknown-method failures when
memory-wikiplugin is off
7. Security Hardening
- Removed busybox/toybox from interpreter safe bin list
- Empty approver list no longer grants explicit approval authorization
- Shell-wrapper detection broadened — blocks
envargument injection - Gateway example credentials in
.env.examplenow fail startup if the placeholder is left in
Upgrade Recommendations
- Customers wanting smarter memory: enable Active Memory
- macOS users: try local voice mode and keep audio off the cloud
- LM Studio users: smoother onboarding
- Everyone: upgrade soon for the security fixes