Twin is an open-source research project that aims to give AI systems continuous, cumulative understanding by correlating events and building reusable situation models, rather than reconstructing context from scratch each conversation. A demonstration shows Claude Sonnet 4.6 using Twin's MCP server to answer questions about a project without any custom memory or local files.
Over the last few months I've realized that I spend an absurd amount of time (and money) teaching the same things to AI over and over again. Information about my projects is already there. Slack contains discussions and decisions. GitHub contains commits and pull requests. Meetings, emails and documents all capture different pieces of the same story. Yet every time I start a new conversation with an LLM, I gather those pieces again and inject them into the prompt so the model can reconstruct an understanding that already existed yesterday. At some point I stopped asking how to retrieve multiple pieces of context and started asking a different question: how can software form, revise and reuse understanding over time? That question led me to start building Twin, an open source engineering research project exploring what happens if AI systems continuously build understanding instead of reconstructing it from scratch every conversation. Most existing projects seems to optimize retrieval, memory or context construction. Twin explores a different layer of the pipeline. It continuously observes distributed events, correlates them, reflects on them and forms situation models that become reusable computational understanding. Instead of giving downstream language models a collection of Slack messages, pull requests or documents and expecting them to connect the dots, Twin tries to do that work beforehand. I recently reached the first milestone that genuinely convinced me this direction might be viable. Using Claude Sonnet 4.6, Twin continuously processed GitHub activity and Slack conversations from a public software project, correlating events and building understanding through reflection over time. After that, I opened a completely fresh Claude conversation. Claude had no custom memory, no project-specific rules, no prompt describing the repository and no access to local project files. The only integration available was Twin's MCP server and automatic context injection. When I asked about the project, Claude didn't receive the Slack messages or the pull requests and infer the situation itself. Twin had already synthesized that understanding. Claude explained why a feature had become a launch blocker, how it had been implemented, which pull request resolved it and how that changed the project's state, even though none of those relationships were explicitly written anywhere. Watching that work for the first time completely changed how I think about AI memory. I don't think the real problem is remembering more anymore. I think it's carrying understanding forward (a.k.a. cognitive continuity). If this idea resonates with you, everything is open source at comments section. I've been thinking about almost nothing else for the past three weeks because I genuinely believe this direction has the potential to change how we build AI systems. The README explains the motivation and research hypotheses in much greater depth, and the repository also includes the complete demonstration shown here, along with additional details and technical context. I'd genuinely appreciate your thoughts, especially if you think I'm wrong.
The article discusses the challenge of maintaining consistent context across multiple AI models like ChatGPT and Claude, presenting three common approaches: manual context passing, using one primary model, and a unified workspace, recommending a hybrid approach anchored to a single source of truth.
This article discusses the limitations of AI models in maintaining context over long conversations, highlighting recency bias and the distinction between context window size and actual comprehension. It suggests practical workarounds like restating constraints and using running context documents.
The author built mem-port, a local MCP server that gives AI copilots (ChatGPT, Claude Code, Cursor, Windsurf) shared long-term memory via embedded SurrealDB, solving context drift between AI tools.
Microsoft claims to have solved the context window problem in AI, addressing the limitation where models must retain every token in their chain-of-thought during complex reasoning tasks.
MIT CSAIL researchers propose a novel approach to AI memory that avoids context rot by storing documents externally and having the AI navigate and query them, achieving 10 million token effective context at lower cost.