Dylan Couzon

AI has gotten better at reasoning, but it still forgets almost everything. Join @DylanCouzon from Qdrant and James Le from @twelve_labs on July 28 for an evening dedicated to AI memory, video intelligence, and retrieval systems. Hear from engineers building production AI systems as they cover: → Persistent memory for video AI → Collective memory for Edge AI with Qdrant Edge → Practical retrieval architectures and live demos Want to showcase your own project? We’re also hosting community demos, and only 2 demo slots remain! If you’re building with AI, retrieval, video intelligence, or agentic workflows, we’d love to have you present your work. 📅 July 28 📍 Bellevue, WA 🎟️ Register here: luma.com/kyksgkak See you there!

My most honest PR

@elastic We measured and @qdrant_engine seems more efficient krabarena.com/claims/qdrant-…

The qdrant-advisor skill is now easily installed to any of your coding assistants: npx skills add qdrant/skills/meta/qdrant-advisor Installing this one skill keeps your agents primed with the latest skills context for your projects and deployments. Even as we continually improve and add to said skills. Read more here: github.com/qdrant/skills

@SakshiSugandhi @themishra4402 Thanks :D

@themishra4402 Qdrant being open source still feels underrated

Popular vibe coding tools and their free alternatives • Cursor → VS Code + Cline • Claude Code → Aider • Windsurf → Continue.dev • Lovable → Bolt.new • v0 → Magic Patterns • Replit AI → Firebase Studio • GitHub Copilot → Codeium • Devin → OpenHands • Augment → Cline • Bolt.new → Webcrumbs • Midjourney → Flux • ElevenLabs → Kokoro TTS • Pinecone → Qdrant • Supabase → Appwrite You don’t need a $200/month AI stack. Most of the best alternatives are open source… and free. Which free tool surprised you the most? 👀

@jreuben1 Shameful title for a piece that spends half its length admitting Qdrant wins on everything except one vendor-supplied throughput number.

I Benchmarked pgvector vs Qdrant vs Pinecone on 50M Vectors — Postgres Crushed the Dedicated DBs by 11x pub.towardsai.net/i-benchmarked-…

Elastic's benchmark claims Qdrant is 7x slower on disk search. They never enabled the two features built for that workload. We turned them on and found 2x throughput, half the latency, 1/3 the compute. Read more: qdrant.tech/blog/benchmark…

@dreyethh @qdrant_engine @kanungle @enochidx @KShivendu_ Thanks for your contribution!

We built 𝐒𝐄𝐍𝐓𝐈𝐍𝐄𝐋 for the 2026 @qdrant_engine “Think Outside the Bot” hackathon Sentinel is a tool that learns what “normal” looks like in a scene and detects anything unusual in real time,it runs directly on a device (no internet needed) and watches camera feeds. Here is how it works: ➜captures camera frames ➜turns them into vector embeddings ➜stores them locally in a memory system (Qdrant) ➜compares new frames to past ones to spot anomalies Features of Sentinel: ➜It’s fully offline (no cloud) ➜no pre training required ➜no manual queries needed ➜learns continuous from what it sees ➜can be updated by simple user gestures Qdrant acts as on the on-device memory enabling fast similarly search with metrics like speed,FPS,memory usage and accuracy tracked live The product is fully live and running,you can check it out here⬇️ 𝐒𝐄𝐍𝐓𝐈𝐍𝐄𝐋 : sentinel-three-steel.vercel.app 𝐆𝐈𝐓𝐇𝐔𝐁 𝐑𝐄𝐏𝐎: github.com/Enoch208/senti… 𝐏𝐀𝐂𝐊𝐀𝐆𝐄 𝐑𝐄𝐋𝐄𝐀𝐒𝐄: github.com/Enoch208/senti… 𝑯𝒆𝒓𝒆 𝒊𝒔 𝒂 3 𝒎𝒊𝒏𝒖𝒕𝒆 𝒗𝒊𝒅𝒆𝒐 𝒘𝒂𝒍𝒌𝒕𝒉𝒓𝒐𝒖𝒈𝒉 𝒐𝒇 𝒉𝒐𝒘 𝒕𝒉𝒆 𝒑𝒓𝒐𝒅𝒖𝒄𝒕 𝒘𝒐𝒓𝒌𝒔 @kanungle @generall931 @KShivendu_

@ClaudioDrews25 Great work, thanks!

Just open-sourced **Memory OS** — a complete hierarchical persistent memory architecture for the Hermes Agent. 🪽 🧠 6 layers, fully local: • Structured facts + trust scoring with feedback loop • Hybrid vector search (Qdrant + BM25) • Self-curating LLM Wiki • Semantic deduplication + surgical context injection Built with love for the Hermes ecosystem. Repo: github.com/ClaudioDrews/m… @NousResearch @Teknium

@jpschroeder How is that different from compiling results from Twitter based on people's sentiment?

We’re announcing: VibeBench, a new benchmark for what actually matters — how models feel when used on real work by experienced software engineers. But, we need your help. Here’s how it works: 1. An initial cohort of 1000 qualified software engineers (join: vibebench.standardagents.ai) 2. Groups of 250 evaluate new models for 2 days on real work. 3. Participants subjectively rank the model relative to other models they have experience with. 4. On day 4 a report is released with objective results derived from the subjective tests. How can you help: 1. We all need this benchmark to exist, but for it to become reality, we need an initial cohort of 1000 qualified software engineers. If that’s you, please join! vibebench.standardagents.ai 2. Repost this! We need to reach as many qualified engineers as we can find. 3. Share this initiative with everyone on your engineering teams. Together we can make this benchmark a reality for all of us.

"Opus 4.7 is our smartest model yet"

Love seeing practical engineering posts. If you’re running @vLLM in production and hitting OOM or unstable load, this guide explains why workload profiling + tuned configs matter more than hardware alone. ai21.com/blog/scaling-v… @AI21Labs

Been messing with RAG pipelines and this actually helped: Forget finding “the perfect chunk size.” Try indexing multiple sizes (100/200/500), retrieving from all, and fusing with RRF. Gains ranged from 1–37% depending on the benchmark. Shoutout @AI21Labs Code 👇

Comparing experiments in Phoenix just got a lot easier. The new List View helps you quickly scan results with per-example metrics, while the Metrics View gives you a high-level look at how changes impact cost, latency, tokens, and more. Upgrade to v11.32.1 to start exploring.

See the new Phoenix evals library in action and bring your questions for our virtual workshop this Thursday! RSVP: luma.com/45eopucf

New feature drop: Label your prompts by use-case, provider, and more!

Excited to see the open source community for @aiDotEngineer Paris! @ArizePhoenix will be out in full force, so if you're there, please stop by and say hi if you're an open source advocate! @aparnadhinak will be giving one of her epic talks on Prompt Learning for Agents, so don't miss that talk @ Tuesday, 3pm on the MainStage À bientôt!

Everyone is shipping agents right now. With so many agent frameworks popping up, the choice comes down to which one actually fits what you want your agent to do. We broke down orchestrator-worker workflows for 6 of the most common frameworks: @Agno, @autogen, @crewai, @openai agents, @langgraph, and @mastra. We then looked at 4 things that matter in practice: 🔹Architecture: how abstractions map to agents 🔹Execution & handoffs: how tasks move between agents 🔹Memory: how information flows (or gets lost) 🔹Error handling: how failures are detected and recovered Along the way we surfaced where the frameworks shine, where they fall short, and what work falls on you as a builder. The analysis shows clear tradeoffs: some frameworks emphasize orchestration over recovery, while others excel at handoffs but leave memory management to the developer. We shared notebooks so you can trace the workflows yourself & see what’s happening under the hood. 👉 arize.com/blog/orchestra… Curious to see how different teams are matching frameworks to their workflows.

💡 Your LLM might ace English queries… but what happens when your users switch to Spanish, Hindi, or Mandarin? For teams building global AI systems, this is the hidden challenge: LLMs often fail to generate correct Cypher queries across languages. That means gaps in reasoning, broken knowledge retrieval, and frustrated users. In our latest blog, Phoenix Ambassador Mohit Talniya shares how to fix this with a multilingual Cypher query evaluation pipeline: 1️⃣ Generate Cypher queries in multiple languages 2️⃣ Evaluate correctness across languages 3️⃣ Trace, analyze, and iterate using Phoenix The payoff: a scalable, standardized way to measure and improve multilingual AI accuracy. 🌍 Don’t let your AI be monolingual. 🔗 Read Mohit’s full breakdown in the comments