MCP server for MyMemory — the user-owned, portable memory vault for AI tools. Give any agent read access to your vault's context and let it propose new memories for your review. One memory for every AI.
Three well-named tools with clear, detailed descriptions (avg 280 chars, well above baseline 194). All tools have explicit input schemas with proper Zod validation and type constraints. Naming follows verb_noun pattern (get_, search_, propose_). Descriptions explain WHAT, WHEN, and consequences. Error handling is present but generic. Output schemas are documented in descriptions but not formally declared. No tool annotations (readOnlyHint/destructiveHint). Missing per-parameter descriptions in schema definitions, Zod descriptions are present but not visible in the JSON schema output that LLMs see.
Fetch the user's MyMemory vault context: a compiled context_block (directives first, then facts/preferences/notes, one per line) plus the raw active entries. Call once at conversation start and apply it — these are the user's standing rules, facts, and preferences across every AI tool they use.
Propose new entries to the user's MyMemory vault when you learn something durable about them. Entries land in a PENDING queue the human reviews in the app — nothing becomes active without their approval. Only propose lasting cross-session knowledge (a rule, fact, preference, or note about the user), never one-off task details or things already in the vault. kinds: directive = a rule ("never …", "always …"), fact, preference, note. Max 20 per call; the server dedupes against existing entries.
Search the user's MyMemory vault for a word or phrase. Case-insensitive substring match over active entries (text and kind). Returns { query, count, matches }.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint). mymemory_get_context and mymemory_search are read-only; mymemory_propose is destructive (writes to pending queue). LLMs cannot infer safety properties without explicit hints.
Output schemas not formally declared. Descriptions mention return structure (context_block + entries JSON, { query, count, matches }, etc.) but no explicit outputSchema in tool registration. LLMs must infer structure from description text.
Error handling returns generic error messages without recovery guidance. E.g., 'No MyMemory API key' is actionable, but HTTP errors (401/403) only surface status + data. Missing: 'Try creating a key at https://www.myclawn.com/mymemory → Keys' for auth failures.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
mymemory_propose lacks confirmation/dry-run pattern. Entries are queued for human review (mitigating factor), but no way to preview what will be proposed before committing. For a write tool, a dry-run mode would improve safety.
mymemory_search implements client-side filtering (pulls full context, filters locally). No pagination or result limit enforcement. At large vault sizes, this could return hundreds of entries, bloating context window. Should cap results at 20-50 and document the limit.