After the v5 launch gave you control over tools and models, the next four releases, shipped between September 20 and 26, went after what Xeref remembers and how it talks to you. The model lineup turned over twice, memory became something you can inspect and pause, and chat learned to ask instead of guess.
v5.9 — Chat Questions You Click
Pick instead of type: when a request is ambiguous or a decision is yours to make, chat now shows a short question with 2–4 options. Click one and your choice goes back into the conversation. To answer something else, just type your reply. It works in the main chat and in project chats.
Google Calendar: your OAuth client secret is now stored encrypted. If you saved one earlier, open Calendar settings and save your credentials again. Connection failures also show Google's own error message and code instead of a generic "access denied".
MCP fixes: "suggest next task" now picks your highest-priority to-do first and skips archived tasks. It used to sort priorities alphabetically, so high-priority tasks came last. Updating a task or note now changes only the fields it is meant to.
v5.8 — Opus 5.5 and a Current Pricing Page
Ultra: Opus 5.5 replaces Opus 5 in the model picker, and Opus Plan Mode now plans with Opus 5.5.
Pro: Grok 4.7 replaces Grok 4.6. The pricing page now lists the current models.
v5.7 — Memory & Context
A real settings page: Settings → Memory & Context lets you turn memory on or off, pause it without deleting anything, allow or archive sensitive topics, and delete all memories. Memory is organised by scope (personal, project, agent) with sensitivity levels and an audit trail.
Memories in chat: relevant personal, project and agent memories are now added to the conversation automatically. You can also manage them from the chat box with /memory remember, /memory forget and /memory list.
Safer saving: passwords, API keys, connection strings and card numbers inside longer text are masked instead of stored, and content that is just a secret is refused. Saving from a workflow webhook or the MCP server now goes through the same safe path as chat. Project and agent memories can no longer leak into unrelated conversations, and webhooks no longer copy raw memory text into usage logs.
Save to Artifacts: ask Xeref to "save this as an artifact" and it lands in your Artifacts library as an editable draft, ready to open, rename or publish. These drafts are grouped under a collapsible "Capabilities drafts" header. To go with it, chat replies are now kept to about 1,500 characters; for longer content Xeref offers to save it to Artifacts instead of pasting it into the chat. The artifacts switch in Settings → Capabilities now works too.
XerefHarness: a new command center under AI Agents for DeepSeek Harness, the open-source agent harness from DeepSeek. Send a message and follow a real run as a timeline of reasoning, tool calls and output, on the model your plan allows. One caveat from the release notes: write tools run as soon as the model calls them, there is no approval step yet, and runs are not saved.
Desktop: Xeref Desktop can now launch when Windows starts.
v5.6 — New Models and XerefJev
Ultra models: GPT-6 Astra, Opus 5, Fable 5.1 and Fugu Ultra v2 are now in the model picker, and Opus Plan Mode plans with Opus 5. Opus 5 is also the default advisor model for Xeref Code direct sessions.
Pro models: GPT Image 2 (text-to-image) joins the list, and the rest were refreshed to their latest versions: Gemini 3.8 Flash, Grok 4.6, GLM 5.3, Kimi K3 and Qwen 3.8 Max. Older versions were replaced, not kept alongside.
XerefJev: a new playground under AI Agents for TypeSafe Jev, a decision model that answers typed Yes/No, Choice and Score questions with probabilities and confidence in about 100 ms. Start from six ready-made use cases (ticket triage, lead scoring, guardrails, expense classification, answer verification, email triage) or build your own, and try the confidence gate that decides what to auto-approve. Available on Pro and Ultra.
The through-line of this stretch: Xeref should be careful with what it keeps and clear about what it needs from you. Memory has switches and masking, long answers go to Artifacts, and ambiguous requests turn into a question you can answer with one click.