SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Meta Launches Muse Wonder Models With Memory and Persistence Features for Agentic Work

0
·1 views

Meta has released Muse Wonder and Muse Wonder Full, a two-model family aimed at agentic tasks, on 12 October 2026, with a 1M-token context window. Its Active Memory feature retains user-relevant details across conversations and resurfaces them when they match a current task, raising questions about where that data is stored and how long it persists. A companion feature, Active Focus, is designed to keep multi-step tasks running through difficulties, retry alternative approaches, and seek confirmation before consequential actions. The model card released alongside it includes capability comparisons against frontier rivals and security evaluation results, with Meta recommending human-in-the-loop safeguards for irreversible actions. Muse Wonder is accessible now via the Meta Model API and Muse Code, while Muse Wonder Full is limited to Muse Code subscribers for the time being.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Open-source MCP server flags malformed Search Console API requests before Google returns silent empty results

Google's Search Console API often returns empty results without any error when a request is invalid, such as an unsupported country code, a non-RE2 regex, or a date range in the future. This makes it hard to distinguish a broken query from a genuinely quiet period, and AI assistants can mistakenly explain the empty output. The developer built open-gsc-mcp, an MIT-licensed local Model Context Protocol server that validates requests before calling the API and reports problems in plain language. The tool also offers change-explanation and indexing-audit features, and it is read-only by default with no telemetry.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Protecting a Customer-Support LLM Agent from PII Leaks With Logging and Masking

A tutorial describes how to secure an LLM-based customer-support assistant for an e-commerce service like Cdiscount, which can look up orders, issue refunds, and send emails. It warns that attackers could use prompt injection or tool abuse to make the agent exfiltrate customer personal data. The proposed defense is ReskPoints, an AI agent logger that records each action with its confidence, parameters, and results. The tool automatically masks sensitive fields such as tokens, passwords, and contact details before logs leave the application. It also supports sampling rules to limit log volume while keeping full visibility into tool calls, and can export traces to platforms like Datadog, Prometheus, or OpenTelemetry.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

206 Sign-in Starts, Zero Signups: Checking Whether AI Visitors Really Convert

A founder saw 206 sign-in starts and no completed signups in a week, which seemed to point to a broken auth flow. The real cause was bot-like traffic: about 190 of 205 Google sign-in clicks in four days came from Germany, many with impossible browser and OS combinations. After filtering that traffic out, weekly visitors fell from 358 to about 85 and sign-in starts fell from 206 to one. The author argues that AI referral traffic should be judged by downstream conversion, starting with removing junk data before computing any rates. The method also covers building an AI-assistant segment using both referrer and utm_source fields, since most ChatGPT visits in one founder's data arrived with no referrer.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

7 CSS Myths Beginners Should Stop Believing

Beginners learning CSS often receive advice that is incomplete or simply wrong, which can lead to confusion and frustration. A common misconception is that developers must memorize every property, when the more practical skill is understanding core concepts and looking up details as needed. Another myth holds that CSS only changes colours, while in reality it governs layout, spacing, typography, sizing, positioning, responsiveness, and animation. The article also argues that good CSS is not defined by using many visual effects, but by applying styles with a clear purpose that supports clarity and usability.

0 comments Read more at DEV Community