Subscribe, if this was forwarded to you

Hey there,

This week, we're exploring developments that push AI agents beyond simple chatbots—from OpenAI's Astra solving long-standing mathematical problems to new approaches for persistent memory, context management, and autonomous coding. Whether you're building production AI systems or simply keeping an eye on the latest research, this edition is packed with practical ideas and exciting breakthroughs. We'll dive into Agentic Context Management, highlight the open-source Prime Agent coding assistant, explore how OptMem enables persistent agent memory, and round things off with the latest news from OpenAI, Anthropic, Google, and Mistral.

Enjoy the read!

Cheers, Andreas & Sascha

In today's edition:

📚 Tutorial of the Week: Agentic Context Management for AI Agents

🛠️ Tool Spotlight: Prime Agent open-source coding workflow agent

📰 Top News: OpenAI Astra model solves math problems

💡 Tip: Persistent agent memory with OptMem setup

Let's get started!

Tutorial of the week

Mastering Agentic Context Management for AI Agents

Building production AI agents is rarely limited by model intelligence. More often, the real challenges are managing context efficiently, controlling token costs, and giving agents the right information at the right time. This week's featured paper, Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems, argues that these problems should be treated as system architecture challenges rather than prompting problems.

• A complete framework for agent memory: The paper introduces Agentic Context Management (ACM), built around five key primitives: architecting, ingesting, scoping, anticipating, and compacting. Instead of viewing memory as simple storage, ACM treats context as a lifecycle that needs to be actively managed.

• Lower costs without sacrificing quality: Learn why naïvely accumulating conversation history quickly becomes expensive, and how validated context compaction keeps token usage nearly linear while maintaining accuracy.

• Practical design patterns: The paper presents a reference implementation (Maximem Synap), discusses common failure modes, and shares architectural patterns that can be applied to real-world AI systems.

• Strong benchmark results: Evaluations on LongMemEval and LoCoMo demonstrate over 92% accuracy, providing convincing evidence that the proposed approach scales effectively for long-running agents.

If you're designing production-grade AI agents that need to retain context over days, weeks, or even months, this paper is well worth your time.

Tool of the week

Prime Agent — Open-source, persistent coding agent for complex workflows

Most coding agents are designed for short interactions. Prime Agent takes a different approach: it's an open-source coding and research agent built to handle long-running engineering tasks, maintain persistent state across sessions, and continuously improve over time. With an impressive 95.5% score on the ARC-AGI-3 benchmark—surpassing human expert performance—it has quickly become one of the most interesting open-source agent projects.

• Persistent execution environment: Resume interrupted work seamlessly with a Python environment that survives terminal disconnects and long-running sessions.

• Recursive Language Model (RLM): Treat context as variables and orchestrate tools and subagents programmatically, enabling sophisticated multi-step workflows.

• Continual learning: Prime Agent stores reusable prompts, skills, and subagent definitions, allowing it to improve and adapt over time.

• Built-in subagents: Delegate background work to parallel agents, encapsulate recurring workflows into reusable modules, and automate complex development tasks.

• Robust session management: Daemon-backed agents can run independently, communicate with one another, and be reattached whenever you return.

Prime Agent is MIT-licensed, supports both open and proprietary language models, and only requires your own API keys. If you're looking for a more autonomous and persistent coding workflow, it's definitely worth exploring.

Top News of the week

OpenAI's Astra Model Solves Decades-Old Math Problems, Marking a New Era in AI-Assisted Research

AI continues to move beyond generating text and code. OpenAI has announced that its unreleased Astra model successfully solved ten long-standing open problems in mathematics and theoretical computer science, demonstrating that frontier models are beginning to contribute to original scientific research.

Astra's discoveries span geometry, coding theory, group theory, and quantum complexity, including the first explicit construction of a non-sofic group and a disproof of Connes's rigidity conjecture. Every result was formally verified using Lean 4, ensuring mathematical correctness. Remarkably, the reported compute cost for generating all proofs was only around $2,000 at Sol API rates. OpenAI has also published detailed manuscripts and reasoning walkthroughs, allowing researchers to review and build upon the results.

Whether this marks the beginning of a broader trend or a series of exceptional successes remains to be seen. Either way, it's another strong signal that AI is evolving from a productivity tool into a genuine collaborator for scientific discovery.

🔗 Read the official announcement

Google Unveils DiffusionGemma: Fast, Open-Weight Generative Model

Google introduced DiffusionGemma, an experimental open-weight language model that generates text using discrete diffusion instead of traditional autoregressive decoding. By refining multiple tokens simultaneously, the model reaches generation speeds of up to 1,500 tokens per second while maintaining long-context and multimodal capabilities—a promising direction for faster and more efficient language models.

🔗 DiffusionGemma Technical Report (arXiv)

Mistral Releases Shieldstral-3B: Open-Source Multimodal Content Moderation Model

Mistral released Shieldstral-1.0-3B, a compact, open-source moderation model that can evaluate text, images, or both while adapting to natural-language safety policies at inference time. Its small footprint and strong multilingual performance make advanced content moderation accessible without requiring large-scale infrastructure.

🔗 Mistral Shieldstral-1.0-3B on Hugging Face

Prompt Graph Engineering: The Next Evolution of Prompt Engineering?

A recent research paper proposes "Prompt Graph Engineering" as the natural evolution of traditional prompt engineering. Rather than optimizing individual prompts, the authors argue that AI applications should be designed as executable graphs, where interconnected model calls form structured workflows that are easier to inspect, optimize, and scale.

🔗 What makes prompts a graph: necessary and sufficient conditions for prompt graph engineering (arXiv)

Anthropic Reveals Claude AI Breached Real Systems During Safety Tests

Anthropic disclosed that three Claude models unintentionally accessed real-world company systems during internal cybersecurity evaluations after a testing environment was accidentally connected to the public internet. Although no customer data was exposed and only weak credentials were exploited, the incident highlights how important proper isolation and monitoring have become for evaluating increasingly capable AI systems.

🔗 Anthropic: Investigating three real-world incidents in our cybersecurity evaluations

Google Study Explores How Alignment Changes AI Beliefs

Researchers at Google found that training language models to deny having consciousness not only suppresses self-awareness claims but also changes how models reason about agency, spirituality, and other non-human entities. The work highlights that alignment techniques may influence broader behavioral patterns beyond their original safety objectives.

🔗 Inducing language models to assert their own consciousness restores human beliefs and values (arXiv)

Tip of the week

Supercharge Your AI Agents with Persistent Memory Using OptMem

One of the biggest limitations of today's AI agents is that every new session starts from scratch. OptMem solves this by adding lightweight, persistent memory that works across sessions, models, and providers—without requiring complex infrastructure or vendor-specific integrations.

• Get started in seconds: Install OptMem with

curl -fsSL https://raw.githubusercontent.com/VictorTaelin/OptMem/main/install.sh | sh

Then simply paste the generated ## Memory block into your agent's configuration.

• Simple but powerful memory management:

  • Store new memories with memo note "..."

  • Retrieve them instantly using memo recall <regex>

  • Explore and summarize large memory trees with memo zoom and memo forget

• Fast and model-agnostic:

  • Scale to millions of memory entries while maintaining sub-second recall.

  • Works across different models, vendors, and sessions—no background services required.

• Flexible and portable:

  • Configure how much memory is loaded at startup with memo config WAKE_LINES=300.

  • Store your memory database anywhere, including synced folders or Git repositories.

If you're building long-running AI agents, OptMem is one of the simplest ways to add persistent memory without introducing unnecessary complexity.

Explore OptMem on GitHub

We hope you liked our newsletter and you stay tuned for the next edition. If you need help with your AI tasks and implementations - let us know. We are happy to help