MCP server that gives your AI writing assistant a constant-size memory of a long document. Chapter 1000 costs the same tokens as chapter 10. Local-only: no account, no API key, no network.
Strong domain-specific tool design with excellent descriptions (50-300 chars, well-optimized for LLM selection). All 6 tools have clear names starting with action verbs (remember_, recall_, list_, check_, entity_, search_). Input schemas are comprehensive with proper JSON Schema types, enums, and constraints. However, output schemas are not explicitly documented in the source code, only inferred from tool behavior. Error handling is minimal; tools lack recovery guidance and actionable error messages. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles (remember_chapter is WRITE, others are READ_ONLY). Parameter descriptions are strong but some lack format constraints (e.g., 'query' in search_memory has no length/pattern guidance).
Find places where the document contradicts its own record: a character who left the story and then acts again, a thread past its deadline. Findings are labelled "certain" (two recorded facts disagree) or "likely" (matched the prose — flashbacks and hearsay are possible, so the offending sentence is quoted for you to judge). It does NOT look for semantic contradictions — that needs a reader, which is you. This covers the part you cannot do: remembering exactly, across a thousand chapters.
Everything on record about one character, place or object: its current state and the chapter-by-chapter timeline of how it got there. Names match case-insensitively.
What the document has promised the reader and not yet delivered, most urgent first, with overdue ones flagged and at most one marked as due now. The [T#] numbering is what remember_chapter matches `ref` against, and it is only valid for the chapter number you pass here — list again if you move on.
THE main tool. Returns one paste-ready block holding everything worth knowing before writing this chapter: entity states, the last few chapter summaries, a sampled skeleton of the whole arc, and older material retrieved as relevant. The block is a constant size no matter how long the document is — chapter 1000 costs the same tokens as chapter 10. Per-section token accounting comes back with it, so when something is missing you can see which section ran out of room rather than guessing. Call this BEFORE drafting, not after.
Output schemas not documented in source code. Tool descriptions explain what is returned (e.g., 'one paste-ready block', 'findings labelled certain/likely'), but formal JSON Schema output definitions are absent. LLMs cannot plan downstream operations without knowing return field names and types.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite clear risk profiles. remember_chapter is destructive (WRITE, modifies state); others are read-only. Annotations enable agents to reason about safety and retry logic.
Minimal error handling and recovery guidance. Tools lack actionable error messages. E.g., if a project name is misspelled, emptyProjectNote() returns a helpful message, but other failure modes (invalid chapter number, malformed entity state) have no documented recovery path.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Fold a finished chapter into the document memory. Before calling, read the chapter and produce: · summary — 2-4 sentences. What CHANGED, not what happened. This is the single biggest lever on recall quality months from now. · entities — every character/place/object whose state moved, with its state as of the end of this chapter. · threads — promises made to the reader. Open new ones; advance or resolve existing ones by their [T#] from list_open_threads. Passing `text` as well stores the prose locally, which is what lets check_continuity quote the exact sentence it tripped on later. Calling again for the same chapter replaces its summary — safe to redo after a revision.
Find chapters by what happened in them, reaching past the recent window. Lexical (BM25) by default and offline. Becomes true semantic search when LONGFORM_MEMORY_EMBED_URL / _KEY / _MODEL are configured — with your own endpoint, not ours. The reply always states which mode actually ran.
search_memory 'query' parameter lacks format constraints. No length limit, pattern, or guidance on query syntax (lexical vs semantic mode). LLMs may pass overly long or malformed queries.
entity_card description is terse (68 chars). While accurate, it lacks context on when to call it vs recall_context. LLMs may conflate the two tools.