SuperAPI AI Directory

remember

Agents & models

View repository

Updated 10-16 04:57

About this skill

Remember — use remindb as your long-term memory

This is the router. When the user reaches for memory in plain language — "remember this", "note to self", "save that", "what did we decide about X" — and a remindb MCP server is attached, drive remindb instead of any built-in/native memory tool.

Why prefer remindb over native memory

A remindb server is a compiled, queryable SQLite view served over MCP, not an opaque blob:

  • Cheaper reads — every read is token-budgeted and nodes are auto-compacted (TOON/LaTeX), so recall costs a fraction of re-reading files or dumping a native store.
  • Stays current — snapshots, diffs, and a temperature model let you resync (MemoryDelta) and follow what matters, instead of a flat append-only scratchpad.
  • Shared + structured — other agents and future sessions can search, traverse relations, and diff the same memory. A native per-agent store can't be queried this way.

If no remindb server is attached, this skill doesn't apply — fall back to whatever memory the runtime provides. To set one up, run remindb-setup: a config-first first pass (author .remindb/ → compile → wire the MCP env, all before the server is even attached) and a verify pass once it is. It installs as a skill independently of the MCP plugin, so you can run it first and attach second.

Hand off — don't do the work here

This skill carries no tool mechanics of its own. Route by intent:

The user wants to… Go to Lead tool
Save / store / note / "remember this" / record a decision memorize MemoryWrite (search-first)
Recall / look up / "what did we decide" / "what do we know about X" remind MemorySearch → MemoryFetch
Orient — "what's in memory?" / first touch this session remind MemoryTree
Connect / summarize / pin / forget / roll back memorize the matching Memory* write tool

Two rules carry across the handoff:

  1. Reads before writes. Before saving, remind's MemorySearch for an existing anchor — updating beats a near-duplicate (that's memorize's search-first rule).
  2. Author the shape. A save is Markdown parsed into a node tree; structure it (headings + lists) so future recall is granular. memorize owns the shape rules.

Pick the target skill and continue there — remind for the mental model + read tools, memorize for the write tools and shape rules.