jevmem Automates Project Memory for Claude Code, Cursor, and Codex

A new open-source utility named jevmem has emerged to address a persistent challenge in AI-assisted software development: maintaining consistent project context across coding sessions. Designed to provide automatic project memory for AI coding agents, the tool currently supports Claude Code, Cursor, Codex, and Claude Desktop. Developed by independent creator Avinash-jetwani and powered by TypeSafe AI, jevmem operates by intercepting conversational turns, evaluating their relevance, and storing concise architectural or decision-based notes directly within the project repository. The system functions through a streamlined pipeline that begins with data scrubbing, proceeds to contextual decision-making, and concludes with memory synthesis. Upon installation, jevmem registers platform-specific hooks or agent rules that continuously monitor user and assistant interactions. Before processing, sensitive information such as API keys, credentials, and financial data patterns are automatically stripped from the conversation stream. The remaining text is then analyzed by the Jev API, which runs a series of probability-based assessments to determine whether the interaction contains a project decision, rule, or bug report. If the value crosses predefined thresholds, a lightweight process extracts a single line of text under two hundred characters. Subsequent updates to the same context automatically tag older entries as superseded, ensuring the memory file remains current without manual cleanup. Independent benchmarking conducted on a held-out set of sixty-six development turns highlights the tool performance advantages. When evaluated against six leading large language models, including GPT-6 Astra, Claude Opus 5.5, and Gemini 3.8 Flash, jevmem achieved a ninety-eight point five percent accuracy rate for save-or-skip decisions and ninety-five point five percent for contextual classification. More notably, the tool processed these decisions in approximately three hundred milliseconds, roughly nine to twelve times faster than competing models that required between two and four seconds. The operational cost per decision stood at roughly one-tenth of a cent, significantly undercutting the pricing of comparable LLM-based solutions. While the top-performing models demonstrated marginal gains in nuanced context tagging, jevmem proved more efficient for high-throughput, real-time memory management where speed and cost efficiency are paramount. Privacy and security architecture remains central to the tool design. All processed data is routed through TypeSafe AI with no telemetry collection, and users can enforce zero-retention protocols compatible with supported gateways. While the scrubbing process effectively removes credential patterns and financial data, names and physical addresses are intentionally excluded from automatic masking, requiring users to consult the provided security documentation for granular control. The utility is fully configurable, allowing developers to adjust decision thresholds, enable verbose logging, and integrate seamlessly into existing version-controlled workflows without polluting repositories with sensitive artifacts. By decoupling memory extraction from heavy inference pipelines, jevmem offers a practical solution for developers seeking persistent context without sacrificing performance or incurring excessive API costs. Its architecture positions it as a lightweight middleware layer that enhances the continuity of AI-driven development workflows across multiple major coding assistants.