Sandra gives your assistant the ability to remember. Whether it actually does is a matter of habit — partly yours, mostly its.
Status: available from Phase 1.
Recall before answering. When asked something about the project, the assistant should look in Sandra before saying "I don't know" or guessing. Most of the value comes from this one habit.
Save as you go. When something durable is decided, it gets recorded — at the time, not in a write-up later that nobody does.
Both are behaviours you ask of your assistant, usually by putting the instruction in your project's
CLAUDE.md (or equivalent). Sandra provides the tools; your instructions decide how they get used.
Good candidates share a property: someone will ask about this again, and the answer is not obvious from the code.
| Worth saving | Why |
|---|---|
| Decisions, with their reasoning | The why is what disappears first, and what people ask for later |
| Contracts between systems | Which side provides what, and who depends on it |
| Conventions the code does not enforce | Naming, structure, "we always do X" |
| Gotchas | The thing that cost someone an afternoon |
| Open questions | So they get resolved rather than rediscovered |
The test: will this still be true and useful in six months? If not, do not save it.
By text — fast, exact, finds what you name. Use it when you know roughly what you are looking for.
By meaning — optional, finds things worded differently from your question. Asking about "authentication" can surface something recorded as "login flow" or written in another language.
Meaning-based search needs one extra container, which runs the model on your own machine — nothing is sent anywhere. Without it, everything else still works. See Questions.
The most useful question is often not "what do we know about X" but "what does X affect".
Because facts are stored as connections, your assistant can start somewhere and walk outward — finding the mobile app is affected by a change nobody described as affecting the mobile app, because something in between connects them.
This is what a graph gives you that a notes file cannot, and it only works if the connections were recorded consistently. Which brings us to the next page.
Next: Why naming matters.