Cursor agents are great at exploring a repo — and expensive at rediscovering the same constraints every session. Decisions live in scrollback. You re-grep the same tree. You re-explain “don’t touch X” for the third time this week.
amem is personal agent memory that stays on your machine. It gives Cursor (and other local hosts) a private, searchable store of durable facts about your work — owners, file anchors, gotchas, decisions — so the next session starts oriented instead of cold.
What it is (plain language)
amem is a local SQLite graph under ~/.amem/ plus a small MCP surface. It is not a cloud RAG product and not a shared company wiki. Your claims stay on the laptop. The installable tool is public; the memory database is not.
How it works in practice
The remember contract is deliberately boring:
amem_context — pull a ranked local packet before broad exploration or a large model prompt
Do the work (still verify anchored files; memory is a map, not source of truth)
amem_remember — write durable outcomes (decisions, constraints, owners, gotchas)
MCP URL shape (UI must be running):
http://127.0.0.1:7843/mcp?workspace=<slug>
For Cursor specifically, amem init --platform cursor installs project guidance / hooks so context can land at session start and prompts, with drafts you approve in the Memory tab. Never store secrets or prompting strategy.
