The 10 fastest growing GitHub repos this week: 1. CloakBrowser (+9.1K stars) Stealth Chromium that passes every bot detection test. Drop-in Playwright replacement with source-level fingerprint patches. 30/30 tests passed. 2. AiToEarn (+4.8K stars) Let's use AI to Earn! 3. agentmemory (+6.9K stars) #1# Persistent memory for AI coding agents based on real-world benchmarks 4. UI-TARS-desktop (+3.5K stars) The Open-Source Multimodal AI Agent Stack: Connecting Cutting-Edge AI Models and Agent Infra 5. 9router (+5.4K stars) Unlimited FREE AI coding. Connect Claude Code, Codex, Cursor, Cline, Copilot, Antigravity to FREE Claude/GPT/Gemini via 40+ providers. Auto-fallback, RTK -40% tokens, never hit limits. 6. DeepSeek-TUI (+8.7K stars) Coding agent for DeepSeek models that runs in your terminal 7. AI-Trader (+3.0K stars) "AI-Trader: 100% Fully-Automated Agent-Native Trading" 8. skills (+18.3K stars) Skills for Real Engineers. Straight from my .claude directory. 9. supersplat (+2.6K stars) 3D Gaussian Splat Editor 10. hysteria (+952 stars) Hysteria is a powerful, lightning fast and censorship resistant proxy. The theme this week: free AI routing hacks and persistent agent memory are the real obsession right now. Bookmark this. Next week's list will look completely different.

Memory Genesis Competition 2026 is in last call — submissions close on March 15. You're also welcome to join us on April 4 at the Computer History Museum for an in-person gathering and high-signal conversations with the EverMind core team and leaders across OpenAI, AWS, research institutes, open-source communities, and the investment world. Guess who will you meet? Follow the competition website for the latest updates: #AIMemory# #AgentMemory# #EverMemOS# #AgenticAI# #Hackathon# #Developers# #AIInfra#

New course: Agent Memory: Building Memory-Aware Agents, built in partnership with @Oracle and taught by @richmondalake and Nacho Martínez. Many agents work well within a single session but their memory resets once the session ends. Consider a research agent working on dozens of papers across multiple days: without memory, it has no way to store and retrieve what it learned across sessions. This short course teaches you to build a memory system that enables agents to persist memory and thereby learn across sessions. You'll design a Memory Manager that handles different memory types, implement semantic tool retrieval that scales without bloating the context, and build write-back pipelines that let your agent autonomously update and refine what it knows over time. Skills you'll gain: - Build persistent memory stores for different agent memory types - Implement a Memory Manager that orchestrates how your agent reads, writes, and retrieves memory - Treat tools as procedural memory and retrieve only relevant ones at inference time using semantic search Join and learn to build agents that remember and improve over time!

🎤 Walrus asked builders at Sui Live whether agent memory should travel with users across platforms and providers. Every answer was yes. Shared context across tools, fewer tokens burned re-explaining yourself, agents that pick up where the last one left off. Memory that lives wherever you signed up first isn't really yours. Data matters. So does who owns it.

We asked builders at Sui Live what would happen if their AI agent lost all of its memory tomorrow. Most said it would be brutal. A few mention having backups. Everyone agreed that starting from zero would be the worst outcome. Agent memory needs to be persistent and portable so that it moves with the user from agent to agent. That's what Walrus is for. 😎

We spent 6 months on one problem: agents losing context in long sessions. Ended up building and open-sourcing an agent memory system. A few things we learned: 🪄compressing stale context mid-session cut token usage by 61% 🪄giving agents a structured task map (mermaid-based) made them way less likely to lose track in 30+ step workflows 🪄persona coherence jumped from 48% to 76% once we added dedicated persona memory repo 👉 Agent memory is genuinely hard and we don't have all the answers. Happy to dig into architecture, benchmarks, tradeoffs, whatever. AMA👇 @TencentDBAbxo2 team is here to talk about it.

New Illinois+ Tsinghua University and other labs study finds that LLM agents still have unreliable memory and that it can get worse when they keep rewriting their own memories. LLM agents can learn from experience, but their rewritten memories often become unreliable. The problem is that many agent systems store past work by asking an LLM to compress messy experience into neat written lessons. That sounds useful because the agent should remember what worked before, but the paper finds that repeated rewriting slowly damages the memory. The core idea is that raw episodes, meaning the actual past attempts and solutions, often stay more useful than the polished lessons made from them. The authors tested this across tasks like web shopping, simulated worlds, app use, and ARC-style puzzle problems where they could control the correct solutions. The sharpest result is that GPT-5.4 solved 100% of a small ARC-AGI set with no memory, but after memory was built from correct solutions, streaming updates dropped it to about 54%. The failures came from bad grouping, overbroad lessons, and overfitting, so the memory forgot details, mixed up task types, or learned rules that only worked on narrow examples. The big deal is that agent memory should not automatically rewrite every experience into a summary, because keeping raw evidence and only sometimes making summaries worked better. The paper is really proposing that agent memory should treat raw past episodes as important evidence, not as disposable notes to summarize away. ---- Paper Link – arxiv. org/abs/2605.12978 Paper Title: "Useful Memories Become Faulty When Continuously Updated by LLMs"

We are entering an extremely exciting era for open-weight models. Kimi K2.6 now feels like a top agentic model. I took it for a spin via @FireworksAI_HQ fast inference APIs. Kimi K2.6 has impressive agentic capabilities, design skills, and the ability to synthesize large amounts of information. I built a little Skill that produces survey papers on any AI research topic you want. (see example in the clip) You can use the skill to tell your agent to generate a survey on whatever topic and watch it go to work. The artifact was fully generated by @Kimi_Moonshot's Kimi K2.6. It's cheap and fast. Next step for me is to explore ways to continue integrating the capabilities of these models on use cases like automating my LLM knowledge bases and augmenting my agent memory capabilities. Stay tuned for more.

“Sui doesn’t have a narrative” “Sui is dead” “Sui is a stablecoin” “Sui is building the best tech stack” “Sui is the next Solana” “Sui is the settlement layer of the future” With the advancements in AI, Sui Move has never been easier to learn. The barrier now is the incentive to move (haha) to a new coding language or start with Move initially. Why you should build with Sui and in the ecosystem: - Hashi allows institutions to put their BTC to work. Earn yield and activate BTC movement directly through the Sui validator network - Sui Dollar, the yield generating stablecoin, allows companies to passively earn rewards on idle cash flows and treasuries - DeepBook is the liquidity layer with Margin, Spot, and now Predict live to build with. Build the best frontends for the primitives - AI agents will need tech that can handle large volume processing and low costs to transact. @b1ackd0g built the Move language with non-human interaction in mind (fast, volume ready, and low cost) - Walrus, Nautilus, and Seal are a prime use case for AI agent memory with access control and privacy - Core components of Move itself making it a secure language. When building in Move, issues with reenterancy are eliminated and token standards are more flexible for your purposes. The object oriented language enables bulk transactions (PTBs) possible. Sui is secure, fast, and easy to use

Creators and project teams today actually own a lot of assets. But most of these assets are only usage rights granted by platforms. Xiaohongshu, Douyin, TikTok, X, YouTube, Claude, OpenAI, Notion, Google Drive, GitHub, Discord. Accounts, followers, content libraries, historical interactions, recommendation weight, API access, payment channels. All of them live under someone else’s rules. TikTok’s official account safety page clearly includes processes for content removal, account bans, appeals, and data downloads. Anthropic’s transparency page also states that policy violations may lead to warnings, suspensions, or termination of access, and disclosed 1.45 million banned accounts in the second half of 2025. This is not a conspiracy theory. This is simply how platform governance works. People felt this less strongly in the past. That made sense. Back then, many people treated internet assets mainly as traffic tools. Losing an account hurt, but it did not always feel like a systemic loss of assets. That has changed. Content, customers, private communities, automated workflows, AI prompts, historical data, agent memory, community relationships, and brand credibility are all now stored online. The more assets accumulate online, the more damaging platform restrictions become. But this problem cannot be solved by a few on-chain platforms alone. Existing assets are already in a vulnerable state. People’s content, followers, transaction records, project reputation, and account weight have long been accumulated inside centralized platforms. Building a new decentralized platform usually does not solve the short-term problem, because users will not automatically migrate, and traffic will not automatically migrate either. A more realistic way to look at this is to break it into four layers. First, the content itself can be made safer. Articles, source video files, images, creative assets, prompts, model outputs, workflows, user research, and community records can all be stored in self-controlled storage, backups, knowledge bases, Git, object storage, or decentralized storage. The goal is simple: if a platform deletes your post or bans your account, you still keep the original assets. Second, identity can be made safer. Account names, domains, wallet addresses, DIDs, email lists, websites, RSS, and newsletters can form an identity layer outside any single platform. Bluesky’s AT Protocol treats account portability as a core design goal, so users can migrate their account if a Personal Data Server fails or stops operating. Nostr also separates identity from any single server through public keys and relays. Third, the social graph can be made partially safer. Follow relationships, subscriptions, address books, community members, and customer lists can be backed up and synced across platforms. But this is much harder, because social relationships have strong network effects. People interact where their habits already are. Exporting the data does not mean the interaction can be exported with it. Fourth, distribution power is extremely hard to decentralize. TikTok’s For You feed, Xiaohongshu’s recommendation system, X timeline, YouTube recommendations, the App Store, and Google Search are all traffic allocation systems. They decide who gets seen. Web3 can preserve your content and identity, but it is very hard to replace the attention-distribution power of centralized recommendation systems. Many Web3 founders die from one illusion: believing that once data is on-chain, users will naturally show up. Reality is heavier than that. Founders have to accept the algorithmic power of TikTok, Xiaohongshu, YouTube, and other major platforms, and accept that social graphs are very hard to make effective across platforms. So the more realistic direction is not to replace every platform. It is to add an escape layer. Centralized platforms can remain the traffic entrance. Your own website, domain, newsletter, private community, content library, wallet identity, and on-chain records become the asset base. Platforms are used for acquisition. The base is used for accumulation. That way, even if one platform goes wrong, your core assets can still be migrated, reused, and redistributed. AI degradation follows a similar logic. Teams should not tie their core production system entirely to one model. A more resilient approach is to keep prompts, workflows, knowledge bases, code, agent configurations, evaluation standards, and historical outputs in places they control. Claude, ChatGPT, Gemini, open-source models, and local models are all just execution layers. Models can change. Core assets and workflows should remain. So the practical strategy is not to fantasize about leaving centralized platforms. Wherever the traffic is, you keep using those platforms. But all core assets should gradually move away from dependency on any single platform. Content needs backups. Identity needs a primary entrance. Users need to be reachable again. Workflows need to be portable. AI production assets need to stay in your own hands. On-chain records should only be used for the most critical states that truly require verification. This is the realistic meaning of Agent Sovereignty. The narrative that AI has a soul, or that AI should own a wallet and make money by itself, is too far away and too likely to attract regulatory pressure. But if Agent Sovereignty means the portability and tamper-resistance of core states, such as memory, permissions, workflows, identity, reputation, and historical behavior records, then it becomes a real need. If a developer spends six months tuning a high-value agent, they absolutely cannot tolerate losing every prompt, output history, and memory because OpenAI or Claude triggers one risk-control action. At the execution level, there are still several traps to watch. First, frictionless experience is the default human preference. Adding an escape layer inevitably adds extra steps. In real life, most people strongly prefer frictionless experiences. If they can take business class on a high-speed train, they do not want to squeeze onto a bus. If they can log in with one click, they do not want to remember a seed phrase. Backups, cross-platform syncing, multisig, and maintaining an on-chain identity are naturally against user behavior. An escape layer only works if the infrastructure becomes extremely smooth. If asset continuity requires creators or developers to spend one extra hour every day maintaining the base layer, the whole solution will collapse. Second, asset portability does not equal asset reusability. A Claude-optimized prompt may produce terrible results when moved to an open-source model. Agent memory accumulated on one platform, such as a JSON file, may not be directly readable by another platform at all. So storage and backup alone are not enough. Real infrastructure also needs to solve standards and formats. Otherwise, what gets exported is unreadable dead data, not live assets that can immediately return to production. Third, only people who have felt the pain are willing to pay. This logic is defensive by nature. Before a systemic crisis happens, ordinary creators and junior developers are unlikely to pay time or money for a probabilistic risk.