The tracking pipeline now automatically extracts relevant entities—such as companies, technologies, and products—from browsing session AI summaries and maps them into a searchable graph. By linking entities that appear within the same session, we can now track relationships and build a structured view of user interests and research patterns. This update significantly enriches the value of historical session data for analysis and discovery.

We've added a suite of new API endpoints to power core features including AI suggestions, autopilot tracking, real-time chat, and graph data visualization. These routes establish the backend foundation for integrating advanced intelligence and interactive analytics into the application. This update effectively enables the connectivity required for our upcoming user-facing AI tools.

Added a new RAG-based chat API alongside semantic document processing utilities to enable intelligent responses based on document understanding. Also introduced a custom-styled Markdown component to ensure clearly rendered and aesthetically pleasing AI-generated content. These changes significantly improve the application's ability to provide context-aware responses to user queries.

This update introduces the foundational dashboard layout and centralized global CSS styles to ensure visual consistency across the application. Additionally, a new MarkdownContent component was implemented to streamline the rendering of formatted documentation and rich text content. These changes provide the necessary UI scaffolding to support upcoming feature development.

The debounceRef in MarkdownContent was missing an explicit initial value, which is safer and clearer when dealing with useRef for timeouts. By initializing it with null, we ensure better type safety and avoid relying on the default undefined value, aligning with modern React and TypeScript best practices.

This update introduces a complete 'Structured Infra' UI overhaul across the entire application, moving to a sharp, monochrome, border-driven aesthetic while deprecating glassmorphism and soft shadows. Alongside the design shift, we've significantly upgraded the chat experience with robust SSE streaming support, debounce logic to prevent UI flickering, and advanced markdown rendering that intelligently formats links, timestamps, and source citations. These changes create a more performant, readable, and highly cohesive environment.

We've overhauled the screenshot functionality to ensure reliable captures while handling complex browser security constraints like tainted canvases and cross-origin image restrictions. By replacing SVG foreignObject rendering with a pure canvas approach and dynamically managing <all_urls> permissions, the extension now provides a stable 'Capture Page' feature directly from the popup. Users will benefit from cleaner, more consistent screenshots without running into privacy-related tainting issues.

To better monitor and optimize the performance of our background screenshot generation, I've added comprehensive timing logs into the screenshot capture flow. By tracking the time taken for canvas initialization, DOM serialization, SVG blob creation, and final image encoding, we can more effectively debug bottlenecks in the image capture pipeline across different page types.

Switched screenshot functionality from the browser's captureVisibleTab API to a more localized canvas-based approach in the content script. This removes the need for broad <all_urls> host permissions, significantly improving the extension's privacy posture and simplifying the manifest requirements.

We've formalized a 'Structured Infra' design system to ensure a consistent, minimal, and high-performance UI across the platform. Alongside this, I've implemented a background service worker using chrome.alarms to handle persistent activity tracking and data batching. These changes provide the foundation for a professional, stable dashboard experience and more efficient background processing.

Introduced a 'Capture Memory + Screenshot' button to the extension overlay, allowing users to manually trigger a data capture and screenshot for their current page. This change updates the content overlay interaction flow to include immediate feedback states (success/failure) for the capture process.

We've refactored the screenshot functionality to capture the active tab at the moment of the request rather than relying on an unreliable tab-switching mechanism. This change ensures that screenshots accurately reflect what the user is currently viewing, preventing issues where the wrong content or previous state was captured. This fix provides a much more stable and predictable screenshot experience.

We removed the Redis lock from the track API route, which was inadvertently causing a bottleneck by triggering 429 rate-limit errors across concurrent batch processes. This change improves throughput and allows track requests to proceed without blocking related batches. The system should now handle concurrent ingestion much more gracefully.

We've upgraded our AI backend from gemini-2.5-flash to the new gemini-3-flash-preview model across all API routes and library utility functions. This migration includes increased token limits in our chat endpoints and expanded context processing for RAG queries, significantly improving the depth and quality of AI-generated responses throughout the platform.

We've overhauled our RAG implementation to provide much richer, more accurate responses. This includes increasing content truncation limits, raising token caps for context compression, expanding retrieval scope (topK and sources), and refining our system prompts for better synthesis and citation. These adjustments ensure the model has access to deeper context and produces more reliable, search-optimized output.

This update introduces a massive suite of productivity-focused features, including a unified global search across memories and notes, a quick-notes overlay with Shadow DOM, and intelligent tab management. We've also streamlined the user experience with a new onboarding wizard, improved context menu integrations, and transitioned to non-blocking toast notifications. These changes, bolstered by new models for Research Rooms and public profiles, represent a significant leap in the platform's core capabilities.

We've upgraded our notification system by implementing rich Sileo toasts across all dashboard pages. These new toasts support descriptive content and adhere to a refined dark theme for a more cohesive UI experience. This update improves feedback clarity and visual consistency throughout the dashboard.

We've added a central Providers wrapper to simplify app state management (Auth, Theme, and Notifications) and introduced a dedicated init-db.js script to automate database configuration. The new script handles schema synchronization via Prisma and, critically, sets up the pgvector extension with the necessary HNSW indexes to support high-performance vector search operations. This streamlines the onboarding process and ensures vector search capability is consistently configured across environments.

This massive release introduces a suite of core features including team memory spaces, a dedicated reading list system, and deep-work focus analytics. It also implements significant infrastructure upgrades, such as a knowledge graph for visualizing browsing research trails, vision-based screenshot analysis using Gemini, and an AI-driven weekly intelligence digest system. These changes dramatically enhance the platform's ability to act as a personal and collaborative knowledge base.

This update brings a massive upgrade to the RAG pipeline, significantly improving how browsing history is retrieved and understood. Text chunking now intelligently respects sentence boundaries instead of cutting off arbitrarily, while multi-query expansion casts a wider search net. The pipeline also introduces LLM-powered reranking to prioritize the best results and dynamically compresses context to keep prompt sizes efficient. Your AI assistant just got much better at connecting the dots in your digital memory! 🧠