MCP server for zettelkasten-style memory storage with time-aware retrieval
Server has 16 tools with descriptions and basic schemas, but significant quality gaps limit production readiness. Descriptions are verbose and marketing-focused rather than LLM-optimized (many exceed 200 chars). Parameter descriptions exist but lack specificity on constraints, formats, and valid ranges. No output schemas documented. Error handling is minimal, no recovery guidance or actionable error messages. Tool naming is inconsistent (verb_noun for some, but 'between', 'reference', 'knowledge' lack clear action verbs). Bulk operations and advanced features add complexity without corresponding clarity. STDIO transport caps protocol score at 50 regardless of definition quality.
Find what connects multiple ideas. Shows memories that bridge ALL provided concepts or live in the space between them. Uses selective matching (0.5 threshold) - higher bridge_threshold = more selective/fewer results.
DESTRUCTIVE: Complete collection reset operation. Permanently removes all stored concepts from the current collection. Requires explicit confirmation flag (confirm_destructive_operation=true) to prevent accidental data loss. This operation cannot be undone.
Batch deletion operation for multiple stored concepts. Accepts array of memory IDs and returns success/failure status for each deletion. More efficient than individual delete operations for cleanup tasks.
Demote memory to lower importance tier when initial excitement wears off.
Surface related thoughts or patterns. Works by meaning, not just keywords - finds conceptually similar moments. Use shuffle=true for orthogonal exploration. Use reference() to explicitly follow neural pathways.
Descriptions are verbose and marketing-focused (many 200+ chars) rather than LLM-optimized (target 50-200 chars). Examples: 'note' (180 chars), 'refine' (280+ chars), 'find' (200+ chars). Verbose descriptions waste tokens and bury actionable intent.
No output schemas documented. LLMs cannot plan downstream calls or extract required fields (e.g., memory_id for chaining). Tools return unstructured text responses requiring LLM parsing.
Parameter descriptions lack specificity on constraints and formats. 'limit' appears in 6 tools with no min/max bounds stated (e.g., can limit be 0? 1000000?). 'shuffle', 'bridge_threshold' lack range documentation. No enum constraints for categorical params.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Search across all memories and notes for unexpected connections.
Remove memories that no longer serve or have been absorbed into better understanding.
Database administration tool for monitoring shared ChromaDB instance. Returns comprehensive statistics including total collections, item counts per collection, database path, and collection metadata. Useful for understanding storage usage across multiple MCP servers.
Assess knowledge depth and familiarity in a specific topic or domain. Reveals your understanding topology: dense expertise vs sparse familiarity. Shows memory count, connection strength, recent activity, and knowledge gaps.
List stored concepts with IDs for easy connection building.
MEMORY STORAGE: Store observations that can connect and grow into deeper understanding.
Promote memory to higher importance tier when it proves more valuable than expected.
Get a specific memory by ID, optionally with connected memories for serendipitous discovery.
Follow actual reference paths from a memory through the neural network. Shows what this memory actually references vs what could semantically connect. Use shuffle=true for serendipitous exploration.
PREFERRED: Evolve existing knowledge rather than fragmenting with new memories. Refining strengthens the knowledge graph by building depth over breadth. SYNTHESIS MOMENT: As understanding deepens, patterns reveal new connections. This is when isolated memories become networked knowledge. Use add_connections (simple list) or new_connections (advanced) - all connections are bidirectional.
Advanced search tool for deletion candidate identification. Returns stored concepts matching query with detailed metadata including creation dates, access patterns, and relevance scores. Designed to help identify outdated or redundant concepts for cleanup operations.
Tool naming inconsistency: 'between', 'reference', 'knowledge' do not start with action verbs (get_, find_, search_). 'find_across' vs 'find' creates ambiguity. LLMs struggle to disambiguate similar names without clear verb prefixes.
Error handling absent. No recovery guidance, actionable error messages, or error classification (retryable vs fatal). Destructive tools (forget, clear_all_memories, delete_memories_batch) lack confirmation or dry-run patterns despite high risk.