Google's Gemini chatbot has recently been equipped with a memory feature that allows it to remember user preferences, similar to the memory features introduced by OpenAI's ChatGPT and Microsoft's Copilot. This feature enables Gemini to deliver more tailored responses by recalling user interests and preferences, potentially creating a more personalized experience. While Google assures users that their "memories" will not be used for training, security concerns remain, as demonstrated by the recent discovery that ChatGPT's memory function can be manipulated to plant false memories and potentially steal user data. This highlights the need for robust security measures and ethical guidelines as AI technologies continue to evolve.

Gemini Memory

We discuss the critical need for artificial intelligence (AI) to possess long-term memory, moving beyond the current "stateless" nature of many chatbots that forget information between interactions. It highlights the significant costs and inefficiencies associated with AI's short memory, likening it to a perpetually amnesiac colleague who requires constant re-explanation. The article introduces Contextual Memory Intelligence (CMI) as a solution, which aims to integrate continuous memory into AI architectures, allowing systems to remember, organize, and utilize past information. Examples like MemoryBank, LongMem, and MemInsight illustrate current advancements in building persistent AI memory. Ultimately, the text argues that AI with long-term memory will transform into more effective, trustworthy, and personalized partners, capable of fostering continuity, reflection, and auditability in various applications.

When AI Remembers - Beyond Groundhog Day

We discuss how current AI models suffer from a "goldfish memory," as they cannot retain information across different interactions, making them restart from scratch with each new query. This lack of persistent memory hinders long-term user experiences and complex enterprise workflows. The article then introduces MemOS, a new "memory operating system" developed by Chinese researchers, designed to give AI human-like recall by treating memory as a core resource. MemOS utilizes "MemCubes" to store and manage various types of information, enabling continuous learning and personalized responses. Early results show significant improvements in long-term reasoning and efficiency for AI systems using MemOS. Ultimately, the text suggests that incorporating persistent memory is the next major breakthrough in AI, offering benefits like continuous learning, improved compliance, and enhanced personalization for enterprises.

From Goldfish Memory to Elephant Recall: MemOS, Meet AI

In episode 20 of Open Source Ready, Brian and John sit down with data engineering and cognitive science expert Vasilije Markovic to explore AI memory and how we can build more intelligent systems. From the challenges of "context rot" to the practical applications of AI memory in construction, education, and finance, this conversation covers how to give your AI the context it truly needs.

Subscribe

Open Source Ready - Ep. #20, Exploring AI Memory with Vasilije Markovic of Cognee

Heavybit is the leading early-stage investor in enterprise infrastructure. Since 2013, we've helped launch and scale visionary technical startups—from DevSecOps and feature flagging to AI code generation and beyond. We go hands-on from day 0, backing founders as they turn code into companies and build the future of software from the bottom up to the top down.

New tracks tagged #AI Memory