TypeSafe AI推出了Jevmem，一款专为Claude Code、Cursor和Codex设计的自动化项目记忆工具 1。该工具能够在代码编辑会话中自动捕获决策、约束、bug和待办事项，并将其保存到JEVMEM.md文件中 1。在后续会话开始时，Jevmem会将相关记忆注入到AI上下文中，帮助模型保持项目连贯性 1。
根据基准测试数据，Jevmem的决策速度相比主流大语言模型快9-12倍，成本低40-60倍，准确率达到98.5%（在保存/跳过决策任务上）和95.5%（在保存并分类决策类型任务上）1。中位决策延迟在API层为0.30秒，端到端延迟为0.6秒 1。该工具支持通过TypeSafe AI API或OpenAI和Anthropic密钥进行处理 1。
Jevmem采用了特殊的历史管理机制，旧决策被标记为[superseded]而非删除，保留完整的决策历史记录 1。现已发布的Jevmem v0.4为早期版本，自动捕获功能目前仅限Claude Code使用 1。需要注意的是，基准测试基于由作者编写的66个保留测试集，并非独立第三方基准 1。
TypeSafe AI has released Jevmem, an automated project memory system designed for Claude Code, Cursor, and Codex that captures and preserves key project information across coding sessions.1 The tool automatically extracts decisions, constraints, bugs, and to-do items from code editing interactions and stores them in a JEVMEM.md file, making this context available to the AI in subsequent work sessions.1
The system maintains a complete decision history by marking superseded entries rather than deleting them, ensuring developers can reference the evolution of project choices.1 According to the developer's benchmarks based on a 66-item test set, Jevmem achieves 98.5 percent accuracy on save/skip decisions and 95.5 percent accuracy when categorizing entry types, while demonstrating median decision latency of 0.30 seconds at the API layer and 0.6 seconds end-to-end.1 The tool operates 9 to 12 times faster than leading large language models at 40 to 60 times lower cost, with pricing comparable to GPT-6 Luna at $0.000089 per operation but delivering tenfold speed improvements.1
Processing can be configured through either the TypeSafe AI API or direct integration with OpenAI and Anthropic keys, with a 2-second timeout budget on the Jev API that triggers automatic session skipping if exceeded.1 Jevmem v0.4 is designated as an early-stage release, with automatic capture functionality currently limited to Claude Code.1
评论
还没有评论，欢迎留下第一条。