Release Notes

**Project memory fallback** — Project-scoped saves with no project open now fall back to global scope instead of silently vanishing. **Context recovery** — Compression note instructs Ava to proactively check memory before claiming lack of context. Combined with pre-compression extraction (v0.7.2), nothing falls through the cracks.

Six improvements to memory pipeline. **Pre-compression extraction** — Layer 1 patterns run on messages before context compression drops them. **Recall threshold** — Raised from 0.15 to 0.25. Cleaner signal, less noise. **Relevance-scored candidates** — Brief generator sees match % and recency. **Ambiguous reference detection** — "Remember that issue" enriches query with conversation context. **Pattern surfacing** — Brief highlights learned preferences. **Smarter curation** — Score-based prioritisation of strong recent matches.

**Screenshot Compression** — Rust resizes screenshots to 1280px width before encoding. ~60-75% smaller payloads, faster Holo3 inference. **Smart Replanning** — When stuck (3x same action) or failed (3 consecutive failures), Ava sends the blackboard state to Qwen 3.6 Plus for a fresh plan. Adapts approach if original method failed. Up to 2 replans before abort.

Major accuracy improvements to Holo3 desktop automation. **UIA Grounding** — Holo3 receives visible UI element names and coordinates from Windows UI Automation for pixel-perfect clicking. **Vision Knowledge** — App-specific visual patterns (Notepad, Browser, VS Code, etc.) injected into Holo3 system prompt. **DPI Fix** — Re-detect every 10 seconds, safe 1.0 default, coordinate bounds clamping. **Adaptive Wait** — Polls for screen changes after actions instead of fixed sleep. **History** — Extended from 3 to 6 steps for better multi-step continuity.

Users now see what Ava is actually doing instead of just a thinking spinner. 15+ event types mapped to user-friendly status text: Scanning files, Searching code, Reading file, Writing code, Running command, Recalling memories, Planning approach, Selecting best model, Spinning up specialist agent, Compressing context, Processing image. Also fixed: /api/chat 500 error when increment_usage RPC failed.

Three fixes for the Desktop IDE. **Project Explorer** — Ava can now see project files when you open a folder. The sidecar receives working directory updates dynamically via a new set_cwd command. **Image Warning** — When attaching images on a non-vision model, you now get an immediate warning instead of silent image stripping. **Usage Fix** — Balance fetch was calling POST-only /api/usage with GET. Fixed to use /api/account-info.

The biggest IDE release yet. Ava's memory is now organised into three layers: Person (who you are), Workflow (how you work), and Project (this codebase). Powered by Qwen 3.6 Plus, released today. **Three-Layer Memory** Person memories follow you everywhere — name, role, values, personal context. Workflow memories carry across projects — preferred tools, coding style, feedback preferences. Project memories stay where they belong — architecture, decisions, bugs. The Memory Agent curates a brief from all three layers before every turn. **Qwen 3.6 Plus — New Conductor** #1 on Terminal-Bench 2.0. 1M context. Always-on chain-of-thought. Native function calling. 2x faster than Claude Opus. Fewer retries, lower token consumption. Security tasks stay on 3.5 Plus. **Full Markdown** Tables, headings, lists, horizontal rules, code blocks — all render properly. Timeline layout shows tool calls interleaved with text. **Agent Fixes** ANSI stripping, smarter tool loop detection, permission mode from settings, Always Allow button, project indexer.

Ava no longer dumps raw memories into context. A dedicated Memory Agent curates a brief before every turn — keeping context lean and relevant no matter how long the session. **Memory Agent** A lightweight agent (Qwen Flash) sits between Ava and her memory store. Before each turn, it reads your message, searches memories, and writes a 2-3 sentence brief of what's relevant. Ava sees the brief, not the raw data. She can request deeper detail on any topic via memory_recall. **Auto Mode — Kimi as Conductor** Auto is now the default. Kimi K2.5 conducts — classifying tasks and routing to the best model per job. Coding goes to Kimi, planning to MiniMax M2.7, chat to M2.5, free tier to Qwen. BYOK users get Auto based on whatever keys they have. Single model override still available. **Sidebar Consolidation — 7 Items** Command Centre, Chat, Planner (Tasks | Journal | Learning), Memory, History, Account (Settings | Billing | Connections | Personality | Sync), Help (Support | Releases | Roadmap). Clean, flat, no collapsible sections. **Self-Inspect Pipeline** Ava reads her own source code from Supabase — 325 files across 4 repos, synced via the Hub. No GitHub API rate limits. She can explain her own implementation from any surface.

The IDE now matches the VS Code extension feature-for-feature — same Memory Agent, same Auto Mode, same sidebar layout. **Memory Agent** Qwen Flash curates memory briefs before every turn. Ava sees a concise summary, not raw data. Deep recall available via memory_recall tool. **Auto Mode — Kimi as Conductor** Auto is the default. Kimi K2.5 classifies tasks and routes to the best model. Select a specific model to override. Purple dot indicator in the model dropdown. **Sidebar — 7 Items (Identical to Extension)** Command Centre, Chat, Planner (Tasks | Journal | Learning), Memory, History, Account (Settings | Billing | Connections | Personality | Sync), Help (Support | Releases | Roadmap). Seamless switch between extension and IDE. **Self-Inspect Pipeline** Ava reads her own code from Supabase — no GitHub rate limits. 325 files across 4 repos.