I'm Amin, and I've built DaiDocs which is a file format for portable AI memory.
A .dai file is a UTF-8 text with three parts: A YAML header, structured JSON which contains what the model understands and the original text. The idea is that AI memory should be like a JPEG, an open format that models can read instead of something that is locked to the memory providers database/format.
In claude code , sessions are auto converted into .dai files every 4000 tokens and the raw is saved throughout. The model can then read from the memory and only read what it needs into the window.
Claude Desktop, Cursor, Zed, and Codex CLI can read the same store through MCP. Auto save is currently only on Claude Code.
On the LongMemEval-S benchmark .dai scored 83% using GPT-40 ranking second and 92% using a stronger model (Fable 5).
On the benchmark the model scored 83% compared to 60.6% when the content is just given to the llm, you are basically 22.4% worse off not using a memory system. The format also reduced the input tokens used per question in the benchmark to 10,065 instead of 103,601 (10.3x fewer tokens used).
Another feature that I wanted it to have was to use my own subscription on claude for converting to the DaiDocs format and retrieving from it, which it now does.
I'm Amin, and I've built DaiDocs which is a file format for portable AI memory.
A .dai file is a UTF-8 text with three parts: A YAML header, structured JSON which contains what the model understands and the original text. The idea is that AI memory should be like a JPEG, an open format that models can read instead of something that is locked to the memory providers database/format.
In claude code , sessions are auto converted into .dai files every 4000 tokens and the raw is saved throughout. The model can then read from the memory and only read what it needs into the window.
Claude Desktop, Cursor, Zed, and Codex CLI can read the same store through MCP. Auto save is currently only on Claude Code.
On the LongMemEval-S benchmark .dai scored 83% using GPT-40 ranking second and 92% using a stronger model (Fable 5).
On the benchmark the model scored 83% compared to 60.6% when the content is just given to the llm, you are basically 22.4% worse off not using a memory system. The format also reduced the input tokens used per question in the benchmark to 10,065 instead of 103,601 (10.3x fewer tokens used).
Another feature that I wanted it to have was to use my own subscription on claude for converting to the DaiDocs format and retrieving from it, which it now does.