Cade

General-purpose AI assistants forget everything between conversations.

Each session starts from nothing, so the same context has to be re-explained before any useful work happens. For anything ongoing, that re-explaining costs more than the answer is worth.

The fix is not a better model. It is giving the assistant a durable place to keep what it learns, and a retrieval discipline for reading it back before it acts.

Treat memory as the product. Cade reads from and writes to a structured Markdown knowledge base with an explicit retrieval protocol, so context accumulates across sessions instead of resetting with each one.

1. 01Knowledge base architecture and retrieval protocol
2. 02Assistant behaviour, boot sequence, and operating rules
3. 03Structured note conventions the system maintains itself
4. 04Local-first design, with no third-party service holding the data

• Memory is plain Markdown, readable and portable without the assistant
• An indexed entry point, so retrieval is a lookup rather than a full scan
• Explicit read-before-write rules, which is what stops the store degrading
• Runs locally against the filesystem, so private context stays private