تم تطوير Digital Scrum Master (DSM) مفتوح المصدر، وهو وكيل ذكاء اصطناعي مستقل قادر على إدارة سير العمل الأجايل بالكامل على Kubernetes. على عكس أدوات إدارة المشاريع المدعومة بالذكاء الاصطناعي الحالية التي تقتصر على توليد التذاكر وتلخيص المهام، يمكن لهذا DSM تنفيذ سبرينتات أجايل من البداية إلى النهاية، مما يعالج فجوة كبيرة في العروض الحالية.

Se ha desarrollado un Digital Scrum Master (DSM) de código abierto, un agente de IA autónomo capaz de gestionar flujos de trabajo ágiles completos en Kubernetes. A diferencia de las herramientas de gestión de proyectos impulsadas por IA existentes, que solo generan tickets y resumen tareas, este DSM puede ejecutar sprints ágiles de principio a fin, abordando una brecha significativa en las ofertas actuales.

Un Digital Scrum Master (DSM) open-source a été développé, un agent IA autonome capable de gérer des flux de travail Agile complets sur Kubernetes. Contrairement aux outils de gestion de projet IA existants qui se contentent de générer des tickets et de résumer des tâches, ce DSM peut exécuter des sprints Agile de bout en bout, comblant ainsi une lacune significative dans les offres actuelles.

An open-source Digital Scrum Master (DSM) has been developed, which is an autonomous AI agent capable of managing complete Agile workflows on Kubernetes. Unlike existing AI project management tools that primarily generate tickets and summarize tasks, this DSM can run Agile sprints end-to-end, addressing a significant gap in current offerings.

Built an AI Agent That Actually Runs Agile Sprints End-to-End (Not Just Ticket Generation)

<p>If you’ve played with AI chatbots or agentic frameworks lately, you’ve probably had the same moment I had - most agents can plan, reason, call tools, run workflows… yet somehow they can’t remember something you said 10 minutes ago.</p>

<p>It’s impressive and frustrating at the same time.</p>

<p>That gap — between advanced reasoning and almost no memory — is quickly becoming one of the biggest things holding AI agents back from feeling genuinely helpful.</p>




<h2>
  
  
  <strong>The Real Bottleneck Isn’t the Model. It’s the Memory.</strong>
</h2>

<p>Most AI chat systems implemented today still live inside the model’s context window. Whatever fits in the prompt is what the agent “knows,” and the moment you step outside that window, it’s gone.</p>

<p>That creates familiar issues:</p>

<ul>
<li>You keep repeating yourself</li>
<li>The agent forgets your preferences</li>
<li>Every new request requires re-explaining</li>
<li>Costs go up because the entire conversation keeps getting re-fed into the model</li>
</ul>

<p>It’s not the model’s fault — this is simply how LLMs work unless the right memory layers are added.</p>




<h2>
  
  
  <strong>Why Memory Matters So Much for Agentic AI</strong>
</h2>

<p>Humans don’t restart every conversation from zero. We use short-term memory to keep track of what’s happening now, and long-term memory to store what matters.</p>

<p>AI agents should work the same way.</p>

<h3>
  
  
  <strong>Short-term memory</strong>
</h3>

<p>Fast, session-level memory used for:</p>

<ul>
<li>The latest user request</li>
<li>The step the agent is currently on</li>
<li>Temporary details needed for task completion</li>
</ul>

<p>Often stored in Redis, in-memory state, or workflow-level context.</p>

<h3>
  
  
  <strong>Long-term memory</strong>
</h3>

<p>Where meaningful information lives:</p>

<ul>
<li>User preferences</li>
<li>Lessons from past conversations</li>
<li>Important facts</li>
<li>Summaries of interactions</li>
</ul>

<p>Stored in vector databases like Pinecone, Weaviate, or Qdrant.</p>

<p>This is what makes an agent <em>feel</em> like it knows you.</p>




<h2>
  
  
  <strong>Good Memory Isn’t Storing Everything — It’s Storing the Right Things</strong>
</h2>

<p>Saving every line of chat is messy and expensive.</p>

<p>Selective, meaningful memory is the key — and summarization makes it possible.</p>

<p>Modern memory systems automatically extract:</p>

<ul>
<li>Key facts</li>
<li>Preferences</li>
<li>Decisions</li>
<li>Context changes</li>
<li>Lessons worth retaining</li>
</ul>

<p>Tools like <strong>Mem0, LlamaIndex Memory, LangGraph memory nodes</strong>, and native memory APIs optimize what to remember and when to recall it.</p>

<p>When this works well, the agent shifts from “chatbot” → “assistant.”</p>




<h2>
  
  
  <strong>The Architecture That Works</strong>
</h2>



<div class="highlight js-code-highlight">
<pre class="highlight plaintext"><code>User Message
     |
     v
[Short-Term Summarizer / Session Memory]
     |
     v
Store summary in Redis / Mem0 (TTL ~4h)
     |
     v
Retrieve relevant memory from:
  - Redis / Mem0 short-term summaries
  - MemU medium-term knowledge
  - Qdrant / Pinecone / Weaviate long-term memories
     |
     v
Memory Orchestrator
     - Performs relevance scoring
     - Resolves conflicts across memory layers
     - Decides what to retrieve for LLM context
     |
     v
LLM Inference (context = only what's relevant)
     |
     v
Response + Memory Creation
     |
     v
Memory Evaluator
     - Scores importance of new info
     - Decides: Short-term / Medium-term / Long-term
     |
     ├─ Yes → Write to Qdrant / MemU (long-term)
     └─ No  → Keep ephemeral (Redis / Mem0)
</code></pre>

</div>






<h2>
  
  
  <strong>What Smarter Memory Unlocks</strong>
</h2>

<p>When agents <em>actually remember</em>, everything changes:</p>

<ul>
<li>No more repetitive conversations</li>
<li>Personalized responses based on your style and goals</li>
<li>Faster interactions with smaller prompts</li>
<li>Agents grow more capable over time</li>
</ul>

<p>Memory transforms AI from reactive to proactive — from chatbot to digital companion.</p>




<h2>
  
  
  <strong>The Real Benefit: Faster, Cheaper, Better AI</strong>
</h2>

<p>Smarter memory doesn’t just improve UX — it improves infrastructure:</p>

<ul>
<li>Fewer tokens sent</li>
<li>Lower inference costs</li>
<li>Less CPU spent on embedding/search</li>
<li>Fewer vector DB operations</li>
</ul>

<p>Agents become more efficient <em>and</em> more human-like.</p>




<h2>
  
  
  <strong>The Bottom Line</strong>
</h2>

<p>Agentic frameworks are powerful — but without memory, even the smartest agent will always feel robotic.</p>

<p>The future of AI agents won’t be defined by:</p>

<ul>
<li>Bigger models</li>
<li>Longer context windows</li>
<li>Higher compute</li>
</ul>

<p>It will be defined by how well agents <strong>remember, learn, and build on past experience.</strong></p>

<p>Once AI agents remember as well as they reason, they won’t just respond better — they’ll understand better.</p>

<p><strong>Memory is what turns a model into a companion.</strong><br>
And when built right—using Redis, Mem0, and vector databases—the system becomes faster, smarter, and more human.</p>

تسلط التطورات الأخيرة في الذكاء الاصطناعي (AI) الضوء على فجوة كبيرة في قدرات الوكلاء الذكيين، خاصة في نقص احتفاظهم بالذاكرة. بينما يمكن لهؤلاء الوكلاء أداء مهام معقدة مثل التخطيط والاستدلال، فإنهم غالبًا ما يفشلون في تذكر التفاعلات السابقة، مما يقلل من فعاليتهم في تقديم المساعدة الشخصية.

Los avances recientes en inteligencia artificial (IA) destacan una brecha significativa en las capacidades de los agentes de IA, especialmente en su falta de retención de memoria. Aunque estos agentes pueden realizar tareas complejas como la planificación y el razonamiento, a menudo no logran recordar interacciones anteriores, lo que disminuye su efectividad para proporcionar asistencia personalizada.

Les avancées récentes en intelligence artificielle (IA) mettent en lumière un écart significatif dans les capacités des agents IA, en particulier leur manque de rétention de mémoire. Bien que ces agents puissent effectuer des tâches complexes telles que la planification et le raisonnement, ils échouent souvent à se souvenir des interactions précédentes, ce qui diminue leur efficacité à fournir une assistance personnalisée.

Recent advancements in artificial intelligence (AI) highlight a significant gap in AI agents' capabilities, particularly their lack of memory retention. While these agents can perform complex tasks such as planning and reasoning, they often fail to remember previous interactions, which diminishes their effectiveness in providing personalized assistance.

It’s All About Memory: The Missing Piece in AI Agents

<img src="https://res.infoq.com/news/2025/11/kubecon-open-source-ai-compute/en/headerimage/kubecon-open-source-ai-header-1764314734603.jpg"/><p>AI workloads are growing more complex in terms of compute and data, and technologies like Kubernetes and PyTorch can help build production-ready AI systems to support them. Robert Nishihara from Anyscale recently spoke at KubeCon + CloudNativeCon North America 2025 Conference about how an AI compute stack comprising Kubernetes, PyTorch, VLLM and Ray technologies can support these new AI workloads.</p> <i>By Srini Penchikala</i>

في مؤتمر KubeCon + CloudNativeCon أمريكا الشمالية 2025، ناقش روبرت نيشihara من Anyscale دمج تقنيات Kubernetes وPyTorch وVLLM وRay لإنشاء مجموعة حسابات AI قوية قادرة على التعامل مع أحمال العمل AI المتزايدة التعقيد. سلطت هذه العرض الضوء على أهمية هذه التقنيات في بناء أنظمة AI جاهزة للإنتاج.

En KubeCon + CloudNativeCon América del Norte 2025, Robert Nishihara de Anyscale habló sobre la integración de las tecnologías Kubernetes, PyTorch, VLLM y Ray para crear una pila de computación AI robusta capaz de manejar cargas de trabajo AI cada vez más complejas. Esta presentación destacó la importancia de estas tecnologías en la construcción de sistemas AI listos para producción.

Lors de KubeCon + CloudNativeCon Amérique du Nord 2025, Robert Nishihara d'Anyscale a discuté de l'intégration des technologies Kubernetes, PyTorch, VLLM et Ray pour créer une pile de calcul AI robuste capable de gérer des charges de travail AI de plus en plus complexes. Cette présentation a souligné l'importance de ces technologies dans la construction de systèmes AI prêts pour la production.

At KubeCon + CloudNativeCon North America 2025, Robert Nishihara from Anyscale discussed the integration of Kubernetes, PyTorch, VLLM, and Ray technologies to create a robust AI compute stack capable of handling increasingly complex AI workloads. This presentation highlighted the importance of these technologies in building production-ready AI systems.

Built an AI Agent That Actually Runs Agile Sprints End-to-End (Not Just Ticket Generation)

Was this article worth reading? Share it

SET.DO | AI Todo List & Planner

Leiga

One More Thing in AI