Закреплённый твит
a trend im seeing amongst teams. if your AI coding assistant keeps "forgetting" your codebase, the fix usually isnt another tool. its telling it where things are.
before you install a "give your agent a memory of your whole codebase" plugins, know this: the people who build AI coding tools already tried that and backed away from it.
early versions of claude code indexed your code into a search database. anthropic ripped that out and did the simple thing instead: let the AI search and read your files the way you would.
sourcegraph builds code search for a living. they tried the index too, then dropped it and went back to plain keyword search.
so treat the index as guilty until proven necessary. you really only need it when BOTH are true:
1) your codebase is huge
2) you keep asking things plain search can't answer, like "what calls this function"
before that, KISS and use what already works. keep a short notes file the AI reads so it knows how your project is laid out. use plain find files for "where does this live."
karpathy does similar for his own knowledge: he keeps it in plain markdown files the AI reads, because the upload all and search setup makes the model relearn everything from scratch every time. he's talking about personal notes, but it's the same move for code.
one real exception: when you don't even know what something is called, a meaning based search helps you find it.
and the promblem with an index, its a point in time snapshot. right after you change code it quietly points the AI at the old version.
so before you add the tool, just ask the AI how your project works.
English
