Welcome back. AI agents can reason, plan, and write code — but long-term memory is still a hard problem. A Silicon Valley startup just cracked it. They hit 99% accuracy on long-term memory by ditching vector databases completely. And they're open-sourcing the whole thing in 11 days.

Also: A senior dev revealed all his Claude Code hacks, ship quality UI with a three-layer agent harness, and why Andrej Karpathy hasn't written code since December.

Supermemory hits 99% accuracy with a new agent memory technique: The AI startup just dropped ASMR (Agentic Search and Memory Retrieval), an experimental system that replaces vector search for parallel teams of specialized agents. These agents actively read, analyze, and extract facts from long conversation histories. This approach boosted their LongMemEval benchmark score from 85% to 98.6% in just a few months, all without a vector database. The team plans to open-source the full code in early April.

Cursor reveals Composer 2 is built on Moonshot's open-source model: The coding startup just admitted that Composer 2 is powered by Moonshot AI's Kimi-k2.5, a 1-trillion-parameter open-source model that went uncredited at launch. Co-founder Aman Sanger acknowledged leaving the Kimi base out of the original announcement. Cursor essentially took that foundation and applied continued pre-training and reinforcement learning using four times the original compute, ultimately pushing the model to frontier-level coding benchmarks.

Musk announces custom chip factory for Tesla, xAI, and SpaceX: The Tesla CEO just unveiled Terafab, an Austin-based semiconductor plant where Tesla and SpaceX will team up to produce proprietary 2-nanometer chips. The facility will manufacture two distinct versions: one designed for edge inference in Tesla vehicles and Optimus robots, and a high-performance variant built for xAI and SpaceX’s space data centers. Musk noted that existing suppliers simply can't keep pace, and the semiconductor industry is moving too slowly for his AI and robotics ambitions.

AI is revolutionizing gaming like never before, but according to Supercell’s CEO Ilkka Paananen, we can’t even fathom what’s next for AI gaming.  

Watch his live interview at the ElevenLabs summit as he explains the philosophy driving Supercell’s massive success with hits like Clash of Clans and Boom Beach, as well as how he’s preparing his company for AI. 

Watch it here.

A protocol backlash is brewing. For months, an increasing number of devs have questioned whether MCP offers real value or simply adds unnecessary overhead to existing APIs. This debate reached a tipping point when Perplexity co-founder and CTO Denis Yarats revealed at the Ask 2026 developer conference that the company is ditching MCP in favor of direct APIs and CLIs.

Not everyone needs a protocol. CLI tools like curl and git already live in model training data, so agents can use them without schema overhead. Charles Chen, an engineering leader at the scheduling startup Motion, argues that a lightweight CLI wrapper is faster and more cost-effective for individual devs who only need to call a few APIs.

But the debate has a blind spot. MCP runs in two modes. Stdio (standard input/output) keeps the server local, on your machine. Remote HTTP puts it on a centralized server that every team connects to. Almost no one in the backlash separates the two. Centralized MCP gives teams one place to manage auth, track tools, and keep prompts consistent across agents. Glean co-founder Tony Gentilcore calls these platform-level questions, the kind CLIs were never built to answer.

Built for the boring layer. The smartest teams aren't picking sides; they’re playing to the strengths of both. They stick with CLIs when a model already understands the tool, but pivot to MCP when they need a unified solution for auth, observability, and governance. The protocol got overhyped early on, but its real story was always enterprise adoption, and that part is just getting started.

If you find yourself repeating the same sequence of prompts across different sessions, you can have Claude Code save that workflow as a reusable slash command.

Meta staff engineer John Kim demonstrated this by asking Claude to pull articles about iOS from Hacker News and save a summary to his local “.claude” directory. Once the task was complete, he simply told Claude:

Claude generated a markdown file in “.claude/skills/” that contains a system prompt capturing the entire workflow. This file is automatically registered as “/fetch-hackernews” command, allowing you to rerun the whole process from scratch. When Kim wanted to add another source, he didn't need to manually edit the file, he simply told Claude:

Claude updated the skill to incorporate the new source. This concept works for any repetitive task: from running test suites to generating changelogs or deployment checklists. Just run through the workflow once, ask Claude to save it, and you’ll have a custom slash command ready for future sessions.

How to master agent teams in Claude Code: This tutorial teaches you how to orchestrate multiple AI sub-agents to handle complex tasks in parallel. It covers practical use cases like full-stack development and debugging, plus best practices for managing context, avoiding file conflicts, and controlling costs.

OpenViking: An open-source context database built for AI agents. It manages everything an agent needs (memory, resources, skills) through a file system paradigm, which means context gets delivered hierarchically and the system can evolve on its own over time.

Automating skill acquisition of agentic repositories: AI models are bad at complex, step-by-step tasks, and teaching them manually takes forever. But if you automatically extract and standardize skills from open-source code instead, learning efficiency jumps 40% while staying at human-level quality.

Grow customers & revenue: Join companies like Google, IBM, and Datadog. Showcase your product to our 200K+ engineers and 100K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team