The Race for AI Memory Is On, and Open-Source Is Entering the Fray
As the capabilities of Large Language Models (LLMs) mature and become increasingly commoditized, the ability to provide AI with coherent, cross-application memory is evolving from a mere product feature into a critical infrastructure for the era of AI Agents. The bottleneck for user experience is no longer the model itself, but the depth of memory and context the model has about the user.
On April 21, 2026, OpenAI released a memory feature called ‘Chronicle’ for its programming assistant, Codex, allowing the AI to recall a user’s work from recent moments or even weeks prior. However, this feature was limited to macOS users with a Pro subscription. Just one day later, on April 22, an open-source project named OpenChronicle appeared on GitHub, offering an equivalent implementation and quickly gaining widespread attention.
Calvin, one of the project leads from Tsinghua University, stated that their team’s core research focuses on Proactive Agents, with memory and context being central to achieving this goal. They are dedicated to exploring what to record and how to record it, in order to make AI better serve all users, especially those who struggle to articulate their needs clearly.
Technical Architecture: A Hybrid, AX Tree-First Approach
To enable AI to understand a user’s actions on their device, a common technical path is screen capture combined with Optical Character Recognition (OCR). However, OpenChronicle opts for a more efficient and precise route: prioritizing the use of the AX Tree.
The AX Tree, or Accessibility Tree, is an interface provided at the macOS system level, originally intended for screen reader software for users with disabilities. It can parse UI elements displayed on the screen—such as application names, window focus, input box text, and button labels—into structured, tree-like text data. Compared to image processing, the AX Tree approach offers three major advantages:
- Low Cost: The computation and token costs of pure text processing are far lower than those of image analysis.
- High Precision: It directly retrieves the text information of UI elements, avoiding potential recognition errors that can occur with OCR.
- Structured Data: The output is clearly structured, making it easy to store, retrieve, and for the model to understand.
Of course, the AX Tree has its limitations. It cannot read the internal content of some applications that use non-native rendering technologies (like Lark or parts of Microsoft Word). Therefore, OpenChronicle employs a hybrid strategy: using the AX Tree as the primary information source, with screen captures triggered by user actions (such as cursor movement, page scrolling, or application switching) serving as a fallback solution for unparsable scenarios. This trigger-based mechanism avoids unnecessary computational overhead and memory clutter in static situations, like when a user is watching a video.
According to the project team’s estimates, the system’s daily cloud model API call cost is about 50 cents under a light workload. For heavy usage scenarios with high-frequency interactions, the cost could rise to 3 to 5 dollars. Notably, users can also opt to connect to locally deployed large models, reducing the cost to merely that of electricity.
Unified Memory Pool: From Application Silos to a Standard Protocol
Most AI products on the market today, such as the built-in memory features of ChatGPT or Claude, operate within their own closed ecosystems, creating ‘memory silos’. When switching between applications or projects, users are forced to manually repeat background information, creating a significant interaction burden.
OpenChronicle’s design philosophy is to return ownership of memory to the user. It builds a unified memory pool on the user’s local device, storing information about all actions across all applications as Markdown files, categorized and flattened across seven dimensions. These dimensions include: information about the user, projects, frequently used tools, topics of interest, contacts, work organizations, and past events.
More critically, the project exposes this local memory pool as a callable ‘tool’ through a standard protocol called MCP (Memory-Context Protocol). This means that any AI Agent or application with Tool-Using capabilities (such as Codex, Claude Desktop, or OpenCode) can connect to this unified memory system with simple configuration. For example, a design decision made by a user in the Cursor code editor can be seamlessly referenced by an AI assistant in another chat window, achieving true cross-application context awareness.
The Ultimate Vision: Laying the Foundation for Proactive Agents
Choosing to open-source OpenChronicle is the project team’s strategy to counter the game-changing moves of industry giants. They believe memory is becoming the next major battleground, with major vendors attempting to lock in users by controlling their context, turning user data into a competitive moat. A local-first, model-agnostic open-source solution ensures that this critical infrastructure serves the user, rather than locking them into a specific platform.
After solving the problems of memory storage and access, the Vida team’s ultimate goal is to build Proactive Agents on top of this infrastructure—agents that can actively serve people. Unlike the current reactive agents that ’don’t act unless asked,' proactive agents can anticipate user needs, offer suggestions, and even execute tasks based on long-term accumulated context.
For example, when a developer returns to their work after an interruption, the agent could proactively restore the previous work environment and thought process. When collaborating across documents, the agent could integrate information from different applications to provide comprehensive insights. This deep understanding of ‘how you work’ will transform AI from a tool into a true intelligent partner, serving everyone and amplifying their creativity.
