🤖 AI Weekly Recap (Week 12)

Happy Sunday! We just had another crazy week in AI. MiniMax just dropped an AI that builds itself while another new open source tool gives your Claude Code infinite memory for free.

And that's not all, here are the most important AI moves you need to know this week.

1. Claude-Mem Gives Claude Code Persistent Memory

One of the most frustrating bottlenecks in AI coding is that your assistant forgets your entire project the second you close the terminal. Enter Claude-Mem: a viral new open-source plugin by developer 'thedotmack' that gives Claude Code persistent, long-term memory across all your sessions.

• It acts as a silent observer, automatically logging every tool execution, bug fix, and architectural decision in the background without manual prompting.

• Instead of hoarding massive raw transcripts, it uses AI to compress these observations into dense, semantic summaries stored locally in a SQLite database.

• It utilizes a "progressive disclosure" retrieval system, feeding Claude a lightweight index of past work and only fetching the deep details when specifically needed.

• By injecting compressed context instead of forcing the AI to re-read your entire codebase every session, it reduces token consumption by up to 95%.

Try it now → github.com/thedotmack/claude-mem

2. MiniMax drops M2.7, AI That Builds Itself

MiniMax has released M2.7, a highly efficient reasoning model that marks a massive shift toward recursive self-improvement in AI. Instead of relying purely on human engineers, MiniMax had the model actively participate in its own development cycle.

• M2.7 autonomously handled 30-50% of the team's reinforcement learning research workflow, including experiment monitoring, debugging, and merge requests.

• The model ran 100+ autonomous loops where it analyzed its own failure trajectories, modified its scaffold code, and ran evals, resulting in a 30% performance improvement on internal benchmarks.

• It proves its capability in the wild, too: over three 24-hour trials on MLE Bench Lite (22 ML competitions), M2.7 trained models that earned a 66.6% medal rate, tying Google's Gemini 3.1.

• It delivers intelligence comparable to GLM-5 but at less than one-third the cost ($0.30 per 1M input / $1.20 output tokens).

Try it now → agent.minimax.io

3. Anthropic Drops "Dispatch" for Remote AI Delegation

Anthropic just launched Dispatch, a new feature inside its Claude Cowork environment that completely changes the AI workflow. Instead of sitting at your computer to prompt a chatbot, Dispatch lets you assign complex tasks to Claude from your phone, go about your day, and come back to finished work on your desktop.

• Dispatch creates a single, persistent conversation thread that syncs seamlessly between the Claude mobile app and your desktop.

• You can text your phone to trigger real actions on your computer, like pulling data from local spreadsheets, searching your Slack, or drafting reports, while you're stuck in transit or grabbing coffee.

• The heavy lifting isn't done in the cloud; all processing happens securely in a local sandbox on your actual desktop machine, using your installed tools and files.

• The catch: Your desktop must remain awake, powered on, and connected to the internet for Claude to actually execute the tasks while you're away.

Try it now → claude.ai/new

4. Google Unveils AI-Led ‘Vibe Designing’ with Stitch

Google has completely revamped Stitch, evolving it from a simple coding companion into a full-blown AI-native software design platform available on Google Labs. Instead of manually pushing pixels or wireframing, you just use natural language prompts to describe the exact feel and purpose of the app you want to build.

• Features a new "infinite canvas" that seamlessly blends text, images, code, and UI components into a single workspace.

• Includes a built-in design agent that understands your overall project context, suggests variations, and tracks different versions so you can experiment without losing progress.

• Introduces voice-based controls, letting you literally talk to the AI to make real-time layout changes or brainstorm alternative interface options.

• Integrates directly with development workflows, allowing you to export UI layouts straight into code through its MCP server, SDK, and AI Studio.

Try it now → stitch.withgoogle.com

5. OpenAI Drops GPT-5.4 Mini & Nano

OpenAI just released the smallest and fastest versions of its GPT-5.4 family, designed to be cheap, hyper-fast workhorses for tasks where massive frontier models are just expensive overkill.

• GPT-5.4 mini is over 2x faster than its predecessor (GPT-5 mini) at coding, reasoning, and tool use, making it perfect for debugging or running as a subagent inside Codex.

• GPT-5.4 nano is an even smaller, lighter model designed purely for high-speed API grunt work like data extraction and classification.

• Mini is rolling out to developers via API, Codex, and ChatGPT (accessible via the "Thinking" feature for Free/Go users), while Nano is strictly API-only.

• This launch is a calculated strike at Anthropic’s current dominance in the AI software engineering market, directly challenging the viral success of Claude Code.

Try N → chatgpt.com

6. Microsoft’s MAI-Image-2 Hits Copilot and Bing

Microsoft AI’s Superintelligence team just released MAI-Image-2, a second-generation text-to-image model built to directly rival Google and OpenAI. It immediately debuted at #3 on the Arena.ai global leaderboard and is already rolling out across Copilot, Bing Image Creator, and the MAI Playground.

• Built with direct input from photographers and visual storytellers, the model heavily prioritizes photorealism, focusing on accurate skin tones, natural lighting, and "lived-in" physical textures.

• It solves the classic AI text problem, reliably generating readable typography for infographics, signage, and posters.

• It specifically targets complex, detailed scene generation, making it highly capable of handling dense cinematic compositions, precise lighting directions, and surreal environments.

• Strategically, this marks a massive shift for Microsoft: by vertically integrating its own top-tier models, it dramatically reduces its long-standing reliance on OpenAI for its core consumer and enterprise products.

Try it now → copilot.microsoft.com