AI assistants forget everything between conversations. You explain your project's conventions on Monday, and on Tuesday you explain them again. You make an architectural decision in one session, and the next session cheerfully contradicts it.
Sandra fixes that. It is a memory your agents write to and read from — decisions, contracts, conventions, gotchas, open questions — that persists across sessions, across machines, and across everyone on the team.
You run it yourself. Nothing leaves your infrastructure unless you explicitly switch on the optional semantic search.
Someone decides on Monday that the mobile app will get its data through a new endpoint rather than the existing one. They mention it to their assistant, which quietly records it.
On Thursday a colleague — different person, different machine, different repository — asks their assistant why the old endpoint is still being called. It looks in Sandra, finds the decision, and says so, including who made it and what it affects.
Nobody wrote a document. Nobody remembered to. The memory accumulated as a side effect of working.
A notes file lets you search text. Sandra lets you ask questions about relationships.
The difference matters when you want things like "every decision that affects the mobile app", or "which contracts does the backend provide that the frontend consumes", or "what superseded this?". Those are questions about how things connect, and a pile of text cannot answer them — but a graph can.
That is the whole idea: Sandra stores facts as connections, not just as sentences.
| Page | |
|---|---|
| How it works | The four ideas Sandra is built from. No code |
| Connecting your agent | Point Claude, or any MCP client, at Sandra |
| Using memory well | Saving and recalling, and what good memory looks like |
| Why naming matters | The one discipline that keeps memory useful as it grows |
| Recipes | Worked examples: team memory, memory shared across repositories |
| Questions | Privacy, limits, backups, cost |
If you want to know how Sandra is built rather than how to use it, see the developer documentation.