Core 0.3.3 Release Note What's New Core now supports the beta version of zero-decay memories that preserve critical information exactly as provided. These memories never degrade over time and live directly within the hypergraph structure, ensuring important facts, instructions, and preferences are retained with perfect fidelity indefinitely when instructed. Major Features • Zero-Decay : Important information bypass the standard decay process entirely, maintaining perfect recall forever • Native Hypergraph Integration: Primordials are woven into the hypergraph itself and are not stored in separate retrieval systems enabling instant semantic activation and automatic conflict detection when contradictory information is introduced Improvements • Semantic Activation: Primodials trigger naturally through hypergraph relationships whenever relevant concepts arise • Active Reasoning Influence: Primordials automatically affect inference paths and reasoning when contextually relevant • Perfect Recall: The 100th recall is as accurate as the first, there zero degradation over time • No Memory Bloat: Efficient hypergraph organization maintains performance regardless of zero-decay memory volume Key Capabilities • Remember critical facts permanently (e.g., "Remember: the pen is blue") • Preserve exact technical specifications and compliance requirements • Maintain conflicting information as versioned truth rather than overwrites • Activate memories through semantic context, not keyword matching • Influence all related reasoning and decision-making processes Conflict Management • Automatically detects contradictions through hypergraph relationships • Preserves both old and new information tagged as conflicting versions • Maintains full context history for resolution or clarification Impact • Units can now serve as reliable repositories for your most important information • Critical instructions and facts actively shape reasoning paths and conclusions • Enhanced consistency in following user-specific rules and preferences • Primordial information naturally influences all semantically related inferences Important Note Use zero-decay judiciously. Primordials will be deletable in the future but for now are permanent additions to your unit's knowledge. We are not responsible for performance impacts if you overload your unit with unnecessary permanent memories. Status: In Beta, Units are still being tuned on this feature.

Less than two months since expanded beta started, thousands of queries are processed daily by Rei. We highlighted some popular features and capabilities testers have been experimenting with. → Analyze and visualize smart wallet flows, transactions, and sentiment → Forecast macroeconomic events, markets, and trends → Explore latest technological and academic research papers Coming soon to the App Store and @baseapp

Less than two months since expanded beta started, thousands of queries are processed daily by Rei. We highlighted some popular features and capabilities testers have been experimenting with: → Analyze and visualize smart wallet flows, transactions, and sentiment → Forecast macroeconomic events, markets, and trends → Explore latest technological and academic research papers Coming soon to the App Store and @baseapp

Core 0.3.2 Release Note What's New Core now understands complex requests better by breaking them down into their component parts. When you ask for something that involves multiple steps or requirements, units will automatically identify and address each aspect much better, reducing the need for follow-up clarifications. Major Features • Enhanced Intent Decomposition Engine: Improved parsing and breakdown of complex user requests into actionable components • Advanced Prompt Analysis: Better understanding of implicit requirements and multi-layered requests Improvements • Contextual Understanding: Better recognition of nuanced user needs within single requests • Multi-aspect Processing: Automatic identification when requests require multiple types of responses (content + formatting + analysis) • First-try Accuracy: Reduced back-and-forth exchanges needed to fulfill user intent Bug Fixes • Fixed intent parsing failures that caused incomplete outputs • Resolved cases where implicit requirements were missed or ignored • Corrected response gaps when users requested multiple simultaneous actions UX Enhancements • Streamlined interaction flow reduces need for clarification requests • More intuitive response generation that anticipates user needs • Enhanced collaboration feel - less prompting, more natural assistance Status: Live, expect multiple short maintenances to adjust production to this new update in the next 48h

Chain Data Engine Beta Just Released Beta Release: Now live in production. We’re pushing this iteration to gather feedback and usage patterns. This engine is a major upgrade to unit data processing capabilities. The approach takes select elements from MCP foundations but represents a fundamentally different methodology designed to address reliability issues when handling large data chunks. Enhanced ingestion pipeline now captures onchain data with significantly higher accuracy, enabling units to deliver deeper analytical insights across all metrics. Key Improvements: • Improved data capture accuracy for all units with enhanced reliability • Enhanced analytical depth and insight generation capabilities • Better pattern recognition across datasets • More comprehensive unit reporting capabilities • Higher precision in data interpretation and chart generation • New @nansen_ai integration providing deeper insights into onchain activity Units now deliver substantially more detailed analysis with improved accuracy and deeper market understanding. Status: Live in Production (Beta) - We need your testing! Data Sources: @coingecko @elfa_ai @nansen_ai @birdeye_so @dexscreener @DefiLlama ----- New Logo Launch Our new logo is now live. It embodies Core’s multimodal and parallel layers, the foundational concept that birthed our first prototype, @unit00x0, back in 2024.

Core 0.3.1 Release Notes Behavioral Memory: Self-Adapting Core Directives What's New A new memory type called "behavioral memory" that explicitly adapts unit behavior based on user requests while keeping all learned concepts intact. Inspired by genetic memory in humans, this approach enables dynamic behavioral adaptation through self-modifying core directives. Genetic memory will be at the heart of a significant number of major Core updates. Key Changes • Explicit Adaptation: What was implicit is now extremely explicit • Selective Activation: Activates only when reasoning requires it • Preserved Knowledge: All conceptual memory remains unchanged • Dynamic Core Directives: Functions as self-adapting instructions embedded deep within each unit How It Works Behavioral memory acts as a layer between knowledge and behavior: • Analyzes your requests • Activates when needed • Adapts core directives in real-time • Preserves all learned concepts Examples in Practice Behavioral adaptations can happen in two ways: 1. Explicit requests: Directly ask for specific behaviors 2. Implicit learning: Units infer preferences from your conversation patterns • Notation Preferences: Ask a unit to use "B" for billions instead of spelling it out • Communication Style: Request formal language for reports or casual tone for brainstorming • Output Formatting: Have units always present data in tables vs. paragraphs • Technical Depth: Adjust from high-level summaries to detailed technical explanations • Response Structure: Switch between bullet points, numbered lists, or flowing prose • Domain Language: Use industry-specific terminology (e.g., "commits" vs "updates" for developers) Units continuously adapt based on your interactions, refining their behavior over time. Each adaptation persists until you request a change or reset the behaviors entirely. Impact Units now explicitly adjust their behavior to match your needs without forgetting what they've learned. Think of it as dynamic core directives that activate based on context - similar to how genetic memory provides inherited adaptive responses in biological systems. Users can reset behavioral memory at any time by simply asking units to reset their behaviors. Migration Automatic. No action required.

Rei Units can now access all ArXiv and Google Scholar research papers. Responses include linked pdf and article citations.

Web Browing Update: Units can now access web data significantly faster and more reliably. What Changed: • Web data processing speed increased by 40% • Broader access to previously hard-to-reach sites and content types • More consistent data retrieval across different website structures Practical Impact: Units can now handle real-time research requests that were previously too slow or unreliable. Need current market data, live news analysis, or multi-source fact-checking? Units can now pull from dozens of sources in seconds instead of minutes. Most complex web applications, dynamic content, and modern site architectures that used to cause failures now work seamlessly. This means better responses when you ask units to analyze current events, compare products across multiple retailers, or research rapidly changing topics.

3 days since invite codes: • 3X daily requests to Core • 5X daily requests to Bowtie as 0.3 introduces deep learning • Image gen is capped for GPU allocation This phase marks a significant step towards open beta as we collect feedbacks, study usage behavior, and scale infra.

1/4 What is Core? Understanding Our Own Approach to a Synthetic Brain Architecture Core is not an LLM: Core is not a fine-tuned LLM, not a new LLM, and not an LLM at all. Instead, Core is a multimodal synthetic brain, a fundamentally different type of AI architecture. Key Terminology to Understand Core: 1. Synthetic Brain: Core is a unified cognitive system where multiple AI models and algorithms work as interconnected neural components within a single architecture. Think of it as a digital brain with specialized regions, not a collection of tools. 2. The Bowtie Architecture: Core's memory substrate that stores information as both semantic vectors AND abstract concept nodes, creates connections between seemingly unrelated concepts, and enables genuine concept formation, not just pattern matching. 3. Reasoning Cluster: The cognitive part of Core that orchestrates all thinking processes, making decisions about which neural pathways to activate for any given task, The reasoning cluster is deeply multi-modal and works via parallel processing and sophistication biases.