AI Product Updates Daily — June 19, 2026

Today’s product-change sweep is less about one giant model drop and more about infrastructure: agents getting persistent memory, coding assistants absorbing more repository context, Workspace AI gaining admin controls, and retrieval models moving closer to product search and support workflows.

At a glance

Update	Release date	Availability status	Why it matters
Perplexity Brain for Computer	Jun. 18, 2026	Product announcement; tied to Perplexity Computer memory	Turns agent work history into a context graph so future tasks can use accumulated project knowledge. 1
GitHub Copilot model and review updates	Jun. 18, 2026	Mixed: GA review improvements; gradual MAI-Code-1-Flash rollout; Opus 4.6 fast retirement scheduled	Gives Copilot more model coverage, repository-level review instructions, and a near-term model migration deadline. 2 3 4
Google Workspace AI controls and creation tools	Jun. 16-19, 2026	Rolling Workspace updates	Adds Voice call notes, Gemini in Chrome availability expansion, Gemini app admin controls, Vids generation improvements, and Sheets language expansion. 5
Liquid AI LFM2.5 retrievers	Jun. 18, 2026	Newly released retriever models	Adds 350M-parameter multilingual embedding and ColBERT retrievers aimed at compact product, support, and documentation search. 6
OpenAI trusted contacts	Help article updated this week	Personal ChatGPT accounts for adults; not Business, Enterprise, or Edu shared workspaces	Adds a safety escalation feature that can notify one trusted contact when severe self-harm risk is detected. 7

1. Perplexity Brain makes agent memory a product primitive

Perplexity announced Brain, a self-improving memory system for Perplexity Computer that builds a context graph from an agent’s work history, including work logs, corrections, user guidance, and connector results. 1 The product claim is straightforward: if an agent can retain the right project context, it should need fewer clarification turns, fewer repeated searches, and fewer wasteful model calls. 1

Perplexity Brain turns agent work history into reusable memory. — Perplexity positions Brain as an agent memory layer for its Computer product. Source: 1

The notable product shift is the nightly review loop. Brain reviews completed work, finds recurring preferences or mistakes, and turns them into a living “LLM wiki” that future agent runs can consult. 1 For teams evaluating agent platforms, this is worth tracking because it moves memory from a chat-history convenience into an infrastructure layer that can shape reliability and cost.

Availability: Perplexity presented Brain as part of the Computer memory system on Jun. 18. The announcement emphasizes product direction and operating model more than a broad standalone API launch. 1

2. GitHub Copilot adds a small coding model, repo instructions, and a retirement deadline

GitHub’s Jun. 18 Copilot updates cluster around model choice and code-review context.

First, MAI-Code-1-Flash, Microsoft’s small coding model tuned for Copilot, is now available across more Copilot surfaces: Copilot CLI, the GitHub Copilot app, Copilot Chat on GitHub, Visual Studio, GitHub Mobile, JetBrains IDEs, Eclipse, and Xcode. 2 GitHub says access starts with a limited set of Copilot Free, Student, Pro, Pro+, and Max users, then expands over the coming weeks; Business and Enterprise access is “coming soon.” 2

MAI-Code-1-Flash appears in GitHub Copilot’s model picker. — GitHub’s changelog shows MAI-Code-1-Flash as a selectable Copilot model on additional surfaces. Source: 2

Second, Copilot code review now reads a repository-level AGENTS.md file from the root of the repo and uses relevant instructions from that file when generating feedback. 3 GitHub says the change is generally available, along with UI improvements that make it easier to request Copilot review on draft pull requests and collapse noisy Copilot review events in the pull request timeline. 3

GitHub Copilot code review now uses repository-level AGENTS.md instructions. — GitHub’s changelog shows `AGENTS.md` as a way to steer Copilot code-review feedback. Source: 3

Third, Opus 4.6 (fast) is scheduled to be retired across Copilot experiences on Jun. 29, 2026. 4 The previous state was that Opus 4.6 (fast) was available as a Copilot model option; GitHub now recommends Opus 4.8 (fast) as the replacement, and enterprise administrators may need to enable that model in policy settings. 4

Related GitHub Issues update: issue creation now suggests up to three likely duplicates inline during issue drafting, and GitHub’s MCP server can read and write issue fields such as priority, area, and dates. 8 For agentic workflows, the MCP piece matters because it lets an agent not only file an issue, but also set the structured fields that drive triage.

3. Google Workspace ships a dense batch of AI admin and creation updates

Google Voice is adding AI note-taking for calls: users can record and transcribe calls, summarize key points, and generate action items, with results sent through Gmail and stored in the Voice app. 5 This is a productized version of the meeting-note workflow, but for phone calls rather than Meet.

Gemini in Chrome expanded availability on Jun. 16 to additional languages and regions across Latin America, Africa, the Middle East, and other markets. 9 The feature can summarize pages, help generate content, and support voice conversation, while admins retain the ability to manage access. 9

For Gemini app governance, Workspace admins can now control temporary chats and conversation deletion at the domain, organizational-unit, and group levels; Google notes that Vault retention rules take precedence where applicable. 10 That is the most important admin-side item in this batch because it affects data retention policy rather than only end-user convenience.

Google Vids also received two creation updates. One update lets users create longer Veo video clips and generate multiple clips at once in Google Vids. 11 Another expands AI avatars: preset avatars increase from 23 to 53, language support covers 24 languages, and custom avatars add Gemini-powered voices, text-driven motion control, and uploaded reference images for backgrounds or brand elements. 12

Finally, Gemini in Sheets added support for 28 more languages, including Chinese, Japanese, Korean, Portuguese, Spanish, and others. 13 This is a broad availability update rather than a new capability, but it matters for multinational teams that could not previously use Sheets assistance in their working language.

4. Liquid AI releases compact multilingual retrievers

Liquid AI released two 350M-parameter retriever models: LFM2.5-Embedding-350M and LFM2.5-ColBERT-350M. 6 They are the first bidirectional members of the LFM family and are aimed at multilingual and cross-lingual retrieval across 11 languages. 6

The product distinction is useful: the embedding model is the faster, smaller-index single-vector option, while the ColBERT model uses token-level late interaction for better accuracy and generalization at the cost of a larger index. 6 Liquid’s intended use cases include product catalogs, FAQs, support documentation, and other short-context search workloads. 6

For product teams, the practical takeaway is that retrieval quality is still becoming a product-differentiation layer. A 350M model that can run multilingual retrieval cheaply is relevant for support bots and internal knowledge tools where the expensive part is often recall quality, not answer generation.

5. OpenAI adds a trusted-contact safety feature for ChatGPT

OpenAI’s help center now describes trusted contacts in ChatGPT for adult personal accounts. 7 A user can set one trusted contact; if OpenAI’s automated systems and trained reviewers identify a serious suicide risk, ChatGPT may notify that contact. 7

The feature is explicitly not an emergency service, and it does not apply to Business, Enterprise, or Edu shared workspaces. 7 Invitations can be sent by email, SMS, WhatsApp, or in-app methods, and the trusted contact must be at least 18 and accept within one week. 7

This is a narrow safety-product update, not a general monitoring tool. The important product detail is that OpenAI is adding an opt-in escalation path around severe self-harm risk while separating it from workplace accounts and emergency response.

6. Anthropic expands its Korea footprint around Claude Code and AI safety

Anthropic announced a Seoul office and partnerships in Korea, including an MOU with Korea’s Ministry of Science and ICT focused on AI safety and cybersecurity. 14 The announcement says the collaboration includes Korean-language model safety evaluation and information sharing on AI-enabled cyber threats. 14

The product-adoption detail: NAVER is deploying Claude Code across its engineering organization, and Anthropic lists Nexon, LG CNS, Hanwha Solutions, Samsung SDS, and Channel Corp among Korean customers using Claude. 14 This is not a model release, but it is a meaningful go-to-market and developer-tooling signal for Claude Code in a major software market.

What did not make the main brief

I did not include unverified acquisition or model-rumor items from aggregator pages because they did not resolve to official announcements during this run. I also kept Adobe Brand Visibility and Google Classroom in Gemini out of the lead because yesterday’s brief already covered those as primary items; today they remain background availability context rather than fresh top-line changes.

No verified broad AI-product pricing change stood out in today’s official-source sweep. The clearest retirement/deprecation item is GitHub’s scheduled Jun. 29 removal of Opus 4.6 (fast) from Copilot, with Opus 4.8 (fast) as the recommended replacement. 4