Memory Distillation MCP Server - extracts and distills memories from conversations
Mnemosyne has 7 tools with complete JSON schemas and descriptions. Naming follows verb_noun patterns (memory_decay, knowledge_attach, search_graph). Descriptions are substantive (100-250 chars), explaining WHAT and WHEN. However, parameter descriptions are inconsistent: some params lack descriptions entirely (e.g., 'connection' in knowledge_attach has no guidance on format), and several descriptions are generic. Error handling is present but minimal, no recovery guidance or actionable error messages visible in the tool definitions. Output schemas are not documented in the tool definitions themselves. Security is reasonable (no secrets in params), but permission gates and audit trails are not evident. Composition is sound: tools are single-purpose and chainable (search_graph → person_key_events). Overall: solid foundation, but parameter documentation and error messaging need strengthening.
Compile an AI conversation into user/agent/derived facts with explicit agent-channel opt-in.
Compile ANY external data source (pasted conversation or raw prose) into the unified knowledge graph as documents, entities, edges and evidence. Use doc_type=dialog with a `messages` array for conversations, or doc_type=text with a `text` string for arbitrary prose. The compiled knowledge is later retrievable to sustain the AI persona.
Register a document (PDF/JSON/TXT/MD) or JSON-backed DB as a searchable adapter; rebuilds the entity linker.
Materialize an attached source into the graph as documents + chapters + evidence.
Deterministic memory decay/forgetting: down-weight and archive stale facts (time/importance/access-frequency). Never deletes facts, so the persona evolution timeline stays reconstructable. High-value persona facts (agent_personality, Relationship) are protected unless force=true. No LLM.
Distill a person's full trajectory into key events: each carries a transparent importance score (evidence richness + participant centrality + turning-point flag) and original-text evidence anchors. The tool supplies evidence only — use it to inform your own personality/turning analysis.
Parameter descriptions incomplete: 'connection' in knowledge_attach, 'entity_links' in knowledge_attach, and 'profile' in generalize_compile lack actionable guidance on format, constraints, or examples.
Output schemas not documented in tool definitions. LLMs cannot infer what fields to expect from memory_decay, knowledge_ingest, or agent_fact_compile responses, forcing them to guess downstream field names.
Error handling lacks recovery guidance. Tool definitions do not specify what errors are retryable, what user actions are needed, or what alternative tools to call on failure.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | <=2025-11-25 | v2 |
Structured knowledge-graph search: find entities by name substring (`query`), object type (`object_type`), or attribute value (`attribute`), optionally scoped to one document (`doc_title`). Returns each match with its document, confidence, relation count, and matched attributes — one call instead of many entity lookups.
knowledge_attach 'entity_links' parameter uses oneOf constraint but lacks clear documentation of when/why to use it. LLMs may omit it or misuse it.
generalize_compile uses oneOf to enforce doc_type-dependent required fields, but the constraint is not explained in parameter descriptions. LLMs may pass both 'messages' and 'text' or neither.