Julien Bordon

Wrote up a thing I kept noticing while building Syncpen: my AI never actually "learned" my project. My notes did and it got to stand on them. The workspace compounds, not the model. #buildinpublic syncpen.io/blog/your-ai-d…

@yujitach I have one question, how do you handle your researches notes accross one or multiple projects? How do you colaborate with other researchers one one specific topic?

最近 Claude Fable を試しているのですが、昨晩ふと、ここ半年ぐらい進展がない共同研究について研究ノートをみせて聞いてみたら、なんと非自明な観察をして、ほぼ解決してしまった。一回目の返事は「計算ミスは一ヶ所みつけましたが同じところで詰まりました」だったが、「そこはこう解決するかなと

@anupamrjp Syncpen, the memory and knowledge base for you and your agent that humans can edit. syncpen.io

Founders, promote ur startup url. Let’s get you some traffic.

@ValencianaAbel Durable is the easy half now — plenty of repos can persist context. Trustworthy is the hard half: can you read what the agent stored, and who approved it? An agent memory you can't open is just a cache with confidence.

THIS REPO GIVES AGENTS MEMORY TencentDB Agent Memory keeps agent memory local and structured. - local-first memory layers - designed for durable context - 123 stars today Agents fail when every session starts from zero. github.com/TencentCloud/T…

@ElitzaVasileva @owndotpage awesome; congratulations!

I crossed $1,000 in revenue in the last 30 days with @owndotpage 🥳 Ironically, this happened during my least productive month, when I I was barely working on the product because of my move. Now it's time to scale. This is just the beginning 🚀

- An agent-memory startup just raised $98M. - Redis shipped a context engine. - LangChain shipped wiki agents. - Meta published on when memory should speak. - All of it is the read path: store, retrieve, extract. The write path ; who approved what the agent wrote, where you fix it when it's wrong, is still uncontested. That's the part I build. syncpen.io

@sekmah "Just add memory" fails twice. First as plumbing — Iris fixes that half. Then as trust: an agent memory nobody reads compounds errors as fast as facts. A memory that compounds unread is a liability with good latency. The second fix isn't infra. It's review.

"Thought 'just add memory' to your AI agent was simple? It became full context engine hell. Redis Iris: the boring fix. #AIAgents #ContextEngine tldr.tech/blog/context-l…

@DataScienceDojo "When memory should talk" assumes the memory is right. The nudge only helps if what's stored is true ; a reminder from a stale record makes the agent confidently wrong. Deciding when memory talks is half the problem. Auditing what it wrote down is the other half.

💡 Most AI agent memory systems answer the wrong question. They focus on what to store and retrieve. Meta's new paper argues the real problem is knowing when to speak up. They call the failure "𝐛𝐞𝐡𝐚𝐯𝐢𝐨𝐫𝐚𝐥 𝐬𝐭𝐚𝐭𝐞 𝐝𝐞𝐜𝐚𝐲": an agent can have the right fact sitting right there in its context and still ignore it two steps later. A requirement gets buried, a failed command gets retried, a diagnosed bug gets treated as new. Here's what they built and found: 🔹 𝐀 𝐬𝐞𝐩𝐚𝐫𝐚𝐭𝐞 𝐦𝐞𝐦𝐨𝐫𝐲 𝐚𝐠𝐞𝐧𝐭 𝐫𝐮𝐧𝐬 𝐚𝐥𝐨𝐧𝐠𝐬𝐢𝐝𝐞 𝐭𝐡𝐞 𝐚𝐜𝐭𝐢𝐨𝐧 𝐚𝐠𝐞𝐧𝐭, watching the trajectory every few steps without touching how the action agent itself works 🔹 𝐈𝐭 𝐡𝐚𝐬 𝐭𝐨 𝐜𝐡𝐨𝐨𝐬𝐞 𝐛𝐞𝐭𝐰𝐞𝐞𝐧 𝐢𝐧𝐣𝐞𝐜𝐭𝐢𝐧𝐠 𝐚 𝐫𝐞𝐦𝐢𝐧𝐝𝐞𝐫 𝐨𝐫 𝐬𝐭𝐚𝐲𝐢𝐧𝐠 𝐬𝐢𝐥𝐞𝐧𝐭. Silence is a real, trained decision, not a default 🔹 𝐒𝐞𝐥𝐞𝐜𝐭𝐢𝐯𝐞 𝐢𝐧𝐭𝐞𝐫𝐯𝐞𝐧𝐭𝐢𝐨𝐧 𝐛𝐞𝐚𝐭 𝐞𝐯𝐞𝐫𝐲 𝐚𝐥𝐭𝐞𝐫𝐧𝐚𝐭𝐢𝐯𝐞 𝐭𝐡𝐞𝐲 𝐭𝐞𝐬𝐭𝐞𝐝, including exposing the full memory bank, always injecting a reminder, advisor-style guidance with no persistent memory, and even Mem0's retrieval layer 🔹 𝐓𝐡𝐞 𝐠𝐚𝐢𝐧𝐬 𝐡𝐞𝐥𝐝 𝐮𝐩 𝐨𝐧 𝐫𝐞𝐚𝐥 𝐛𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤𝐬: Claude Sonnet 4.5 jumped from 37.6% to 45.9% on Terminal-Bench and from 55.0% to 61.8% on tau-2-Bench, and even Opus 4.6 still picked up 2 to 3 points The mechanism is simple: knowing something and acting on it are different problems. A model can technically have a fact in context and still fail to apply it at the moment that matters. This memory agent's whole job is catching that gap and closing it with a short, targeted nudge. If you're building long-running agents, the lesson isn't "𝐚𝐝𝐝 𝐦𝐨𝐫𝐞 𝐦𝐞𝐦𝐨𝐫𝐲." It's "𝐝𝐞𝐜𝐢𝐝𝐞 𝐰𝐡𝐞𝐧 𝐦𝐞𝐦𝐨𝐫𝐲 𝐬𝐡𝐨𝐮𝐥𝐝 𝐭𝐚𝐥𝐤." #AI #LLMAgents #AgenticAI #AIEngineering #LLM #MetaAI #AIResearch

@maarcoofdezz The five roles cover the work. Nobody covers the memory. All that markdown the agents write is the system's real state ; and if no one reviews those writes, it drifts into fiction quietly. The skill isn't just directing agents. It's signing off on what they write down.

Así escala de verdad un Senior Engineer con Claude Code. La diferencia está en mover tu tiempo hacia lo que más importa: → mejores prompts, más planificación, más review, menos tecleo. El workflow: usa un plugin que divide cada tarea entre 5 agentes: - uno hace brainstorming - otro diseña el plan técnico - otro implementa - otro revisa - otro valida distintos ángulos del proyecto Todo queda documentado en markdown. Es más lento, hay más espera, pero la calidad sube porque cada agente tiene un rol claro. Y luego viene el multiplicador real: → git worktrees Si Claude Code ya te hace más rápido, los worktrees te dejan lanzar varias sesiones en paralelo, cada una trabajando en una tarea distinta. Su equipo corre 4–8 sesiones de Claude Code a la vez. La nueva skill de ingeniería no es solo saber codear. Es saber dirigir varios agentes sin perder el control del sistema.

@karpathy A teammate is someone whose work you can review. That's the piece this paradigm still needs: the entity's memory as something the org can read; who wrote what, who approved it. Slack is where it talks; your LLM-wiki is where it should remember.

This is a new paradigm for interacting with Claude that is significantly more "inline" with all the other human activity org-wide. Once you do all of the under the hood engineering work to make this "just work" (e.g. across tools, integrations, compute environments, memory, security, etc.), Claude basically joins the team in a seamless way - you can talk to it as you would talk to a person and it can help with a very large variety of workloads. Imo this is the 3rd major redesign of LLM UIUX. The first paradigm was that the LLM is a website you go to, the second was that it is an app you download to your computer. This third one is that it is a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans. It really takes a while to wrap your head around it, but it works and it is awesome.

@T_Reaves14 @BraceSproul @hwchase17 @devstein64 @jeffreyhuber That's the best outcome a paper can have; it changed our roadmap the same day. If you migrate your fleet through Syncpen I'd genuinely value it: you'd be testing the write-path fixes against the protocol that motivated them. Happy to help you set up ; DMs open.

@getsyncpen @BraceSproul @hwchase17 @devstein64 @jeffreyhuber Thank you—this is exactly the kind of response I hoped for. You didn’t just read it; you tested the protocol against a real system and found a failure mode I hadn’t anticipated. I’m also realizing Syncpen could be a real help as I migrate my own agents to Soft-Cache.

1/ I watched LangChain’s webinar on LLM wikis and agent memory and wrote up a working paper from my own small agent-fleet experiments: Soft-Cache: A Human-Supervisable Coherence Protocol for Persistent Agent Fleets github.com/treaves-GSD/so…

@georgerowan0 @X Building AI memory where humans and agents can edit!

Looking to connect people on @X if you're into - design - building SaaS - vibe coding - AI tools - shipping in public - figuring it out as you go let's talk 👇

@davejoh Hi there~ Just read your article on Forbes The Best AI Writing Tools. Could you have a try to syncpen.io ?

@jeffreyhuber If context is the currency, staleness is the inflation. An agent writing to its own memory unreviewed is printing money ; it spends fine until reality audits it. Currencies hold value because someone verifies the mint.

context is the currency of agent performance

I tested your protocol against Syncpen, where humans and agents share a workspace. We pass the identity half: stable doc IDs, agent edits as signed suggestions a human approves. We fail the write path: an agent can overwrite a doc without checking anything first. Fixing that now.

4/ The line I keep coming back to: A cache can agree perfectly with its derivation chain and still be wrong if nobody checked whether reality ever matched the record. Critique welcome. Thanks to @BraceSproul @hwchase17 @devstein64 @jeffreyhuber for the webinar framing.

@buildwbhoomika thanks for the Follow. All the nbest to Sentorix ~!

@launch_llama Building Syncpen — a writing workspace your AI agents can actually write in, not just chat about. Markdown, real-time, every agent edit signed so you stay the editor. syncpen.io

Hey all! Looking to #connect with people who are: - Data engineers turned founders - Developers turned builders - Employees turned entrepreneurs Sound like you? Drop what you're building 👇

@lb_ASC Thank you ! Laurent.

@getsyncpen Bravo Julien, big fan of Syncpen

Wrote up a thing I kept noticing while building Syncpen: my AI never actually "learned" my project. My notes did and it got to stand on them. The workspace compounds, not the model. #buildinpublic syncpen.io/blog/your-ai-d…

Agree folder-per-project is the right shape. The catch with a local folder: it's a place the agent stands, but you can't stand there with it. The moment you want to read what it wrote, fix a decision, or hand the context to a teammate, a local folder is the wrong container. "Coworker" only lands if the human can work the same folder — not just the agent

ONE FOLDER PER PROJECT IS THE SIMPLEST WAY TO GIVE AI WORKING MEMORY. Most people ask Claude to help with everything from one messy chat. Personal finances, home improvement, a new website, and a client project all get blended into the same context soup. The better setup is boring: one Cowork OS folder, then one folder per project. Each folder holds the notes, files, decisions, links, session logs, and next actions for that specific domain. The detail most people miss: AI does not need a mystical second brain first. It needs a clean place to stand before it starts working. When Claude opens the client folder, it should only see the client context. When it opens the website folder, it should inherit the strategy, assets, tasks, and last session notes for that project. That is when the model stops acting like a generic chatbot. It becomes a coworker with local memory, clean boundaries, and a real workflow around the work.