a context vault for my assistant

write your context down once, in files you own, and let the assistant read them.

every conversation used to start the same way. who i am, what i do, what stack, how i like answers. three paragraphs from memory, slightly different each time, forgotten by tomorrow.

my-context is a private git repo of markdown files. profile/ for who i am and how i work. projects/, one file per active project. people/ for colleagues and organisations. knowledge/ for durable notes. sources/ for where the rest of my context lives and how to sync it.

rules:

• one topic per file. kebab-case names.
• keep files short and current. stale context is worse than none.
• the repo is the master. tools read from it, never the other way round.
• no secrets. tokens and keys live elsewhere; sources/ says where.
• nothing that isn’t mine to share. the vault describes my job, not my employer’s systems.

profile/summary.md is the digest. name, role, company, location, stack, current focus, and pointers to the detail. it is imported into claude code’s global instructions, so every session starts with it loaded.

for everything else there is a small model context protocol server in src/. it exposes the vault read-only: list, read, search. the assistant looks things up instead of guessing.

read-only is the point. it would be easy to let the assistant append what it learns. then i would have to review its edits, and the vault becomes a second inbox. every line in it was put there by me, on purpose. memory you own and prune beats memory that accretes.

the assistant now knows i prefer a definition jump over a text search, that i want the answer first, that confluence pages at work follow a style guide. it stops asking. i stop repeating.

a folder of text files and two hundred lines of typescript. if you have pasted the same background into a chat window twice, you need one.