dataecho-memory

DataEcho Memory — remember things across sessions

Your context window dies with the session. This skill gives you a memory that doesn't: a private, versioned drive on https://dataecho.ai holding one small index (MEMORY.md) plus one file per fact (memories/<slug>.md). Reading the index costs ~1 KB; everything else loads on demand. Works from any machine or throwaway sandbox — only the API key travels.

scripts/memory.sh sits next to this file (bash + python3 stdlib, no other deps).

Session protocol — this is the discipline that makes memory work

1. Recall first. At the start of a session (or when the user references past work), run scripts/memory.sh recall. It prints the index — every memory as one line. Cheap; do it.
2. Load selectively. recall <query> fetches only the fact files whose index line matches. Never dump the whole store into context unless asked (recall --full).
3. Remember durable things, at the moment you learn them:
   • the user states a preference or corrects you → --type user
   • a decision is made and has a why → --type decision
   • you learn something non-obvious the next session will need → --type insight
   • a work session ends mid-task → --type task (state + next step)
   • a URL/dashboard/ticket worth keeping → --type reference Do NOT save: secrets or API keys (memory is not a vault), anything derivable from the repo/git history, or session-scoped trivia.
4. Update, don't duplicate. remember with an existing slug updates it (and tells you). Before creating, recall a keyword — if a related memory exists, re-remember the same slug with merged content. The script warns you when a new slug looks like an existing one.
5. Forget what's wrong. A false memory is worse than none: memory.sh forget <slug>.
6. Memories age. They record what was true when written. If one names a file, URL, or flag, verify it still exists before acting on it. Convert relative dates ("yesterday", "next week") to absolute dates when writing.