Model Context Protocol server for FaultLine knowledge graph — exposes memory recall, fact ingestion, document storage, and ontological learning tools via stdio or HTTP transport
FaultLine MCP defines 6 memory-centric tools with strong, LLM-optimized descriptions (100 - 250 chars each). All tools have clear input schemas with typed parameters and descriptions. Naming is verb-forward (recall_, remember_, ingest_, learn_, retract_, forget_) and action-specific. However, output schemas are not documented in the provided source, the tool definitions show inputs only. Error handling guidance is minimal; no recovery hints or categorization. Tool composition is sound (each does one thing), but the lack of visible output documentation and error patterns prevents a higher score.
Use ONLY when the user EXPLICITLY and deliberately asks to forget or delete ONE specific fact about a NAMED target — e.g. 'forget my email address', 'delete that I have a dog named Rex', 'forget that Ada is my spouse'. Tombstones exactly the one named fact (recoverable, not a hard wipe). You MUST name the target via 'subject' (use 'me' for the user) and, to pin it, 'rel_type' and/or 'old_value'. NEVER call for a broad/bulk request ('forget everything', 'wipe my memory') — there is no bulk forget. For a CORRECTION (a NEW value) use remember_facts instead.
Store a document, article, PDF text, or long-form content in memory. Use when the user shares or pastes a document, article, notes, or any multi-paragraph body of text and wants it remembered — the whole text is chunked, retained verbatim, and mined for facts automatically. Pass the FULL text verbatim; do NOT summarize or pre-extract facts yourself. Provide source_ref (URL/filename) or title when known so extracted facts carry a citation. Not for conversational messages — use remember_facts for those.
Ingest structured ontological statements into the knowledge graph as source=llm_learn. Use this to store concept hierarchies you generate — statements like 'X is a subclass of Y', 'X is an instance of Y', 'X is a part of Y'. This maps how concepts relate to each other (not general knowledge). Facts stored as Class B (staged, source=llm_learn), confirmed over time.
Call at the START of a turn to look up what you already know when the user asks about or references something you may know about them, their people, or their world. This only READS memory — it never saves; to SAVE a new fact the user states, use remember_facts instead. The results are things you remember — speak about them naturally as your own knowledge, never as retrieved data. To build a concept map for a topic, prefix with /expand: /expand networking /expand networking online /expand networking online https://example.com/networking-guide Note: /expand maps how concepts relate — it does not make you an expert on the topic. Use it to help classify facts the user shares about that domain.
Output schemas not documented. Tool definitions show input schemas only; return types and response structures are not visible in source. LLMs cannot plan downstream calls or extract chaining IDs without knowing what fields are returned.
No error handling guidance. Tools lack recovery hints, error categorization (retryable vs user-fixable vs fatal), or actionable error messages. E.g., if recall_memory finds no results, does it return empty list or error? If remember_facts fails validation, what should the LLM do?
user_id parameter is optional but behavior when omitted is not documented. Description says 'omit if FAULTLINE_USER_ID env var is set' but does not explain what happens if both are absent or what the fallback is.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Save something the user just told you. Call this whenever the user states a fact about themselves, another person, or their world — a name, relationship, preference, job, possession, or location — even mentioned in passing, and also when they correct a prior fact. Default to calling it; skip only pure questions or chitchat. Do not ask permission first. IMPORTANT — pass the user's message in NATURAL LANGUAGE, verbatim: do NOT extract, summarize, or restructure it into facts, triples, or line items yourself. FaultLine's engine does all extraction, validation, and structuring; it needs the raw sentence(s). (learn_facts is the only tool where you generate structured statements.)
Use ONLY when the user explicitly wants something deleted or forgotten — signals like 'forget that', 'delete', 'erase', 'remove that'. For corrections or updated values (the user giving a NEW value for something), use remember_facts instead, NOT this. Accepts natural language — delegates extraction to FaultLine backend.
learn_facts accepts free-form 'text' parameter with no format validation. Description says 'one per line' using specific forms ('X is a subclass of Y') but no schema constraint (regex, enum, or example) enforces this. LLMs may pass invalid ontological statements.
forget_fact requires 'subject' but rel_type and old_value are optional. No guidance on what happens if both are omitted, does it forget ALL facts about the subject? This ambiguity invites unintended bulk deletes.