OpenAI Codex Head Tibo Says Better Memory Can Shorten Prompts and Enhance Token Utility
Tibo, head of OpenAI Codex and ChatGPT, pointed out that a better memory system equals shorter prompts, leading to higher utility per token.
This view echoes OpenAI's recently launched enhanced memory feature for ChatGPT, which retains context across conversations and maintains usefulness over time, reducing the need for repeated inputs.
In market mechanisms, developers and agent users buy efficient memory and context management tools, selling lengthy prompts and high token consumption; the event-driven iteration of OpenAI's memory function directs funds toward context optimization models and agent development platforms, benefiting OpenAI and users relying on long-context efficiency, while putting pressure on tools and services that depend solely on stateless prompts.
Source: Public Information
ABAB AI Insight
OpenAI has iterated on the ChatGPT memory function multiple times; Tibo's statement based on the new memory system continues the Codex team's long-term focus on optimizing developer tool efficiency, from early code completion to current cross-conversation context management, with the core aim of reducing inference costs.
From a capital perspective, OpenAI is mobilizing user and agent token budgets toward high-value tasks by enhancing memory capabilities, motivated by the desire to improve product stickiness and practicality, strategically differentiating memory as a core infrastructure, and accelerating the evolution of agents from experimental to production-level deployment.
Similar to platforms like Hugging Face emphasizing the observation of abstract layers saving tokens, OpenAI is currently in a phase of transforming AI from prompt engineering dominance to system-level memory and state management, with memory optimization becoming a competitive focus.
Essentially, this is a technological substitution: advanced memory systems replace repetitive prompt inputs with persistent context, directly compressing token consumption and enhancing output consistency, driving AI development from zero-start in each conversation to cumulative intelligent state reconstruction, amplifying the marginal utility of each token.
ABAB News · Cognitive Law
Memory is a compression lever; short prompts outperform long reasoning.
Context accumulation determines token efficiency; wise ones store first and use strong later.
In the agent era, the core cost lies in memory; whoever optimizes state controls practical pricing power.