Persistent long-term memory via MCP — a brain-like persistent storage system for Claude with semantic search, note classification, and vault indexing
GYSTC has 10 tools with complete input schemas and descriptions, but quality is uneven. Naming follows verb_noun convention (brain_store, brain_retrieve, brain_classify) which is good. Descriptions are present but often generic (e.g., brain_regions: 'List all 12 brain regions and their descriptions', 54 chars, below the 100-char baseline for discovery tools). Parameters have types and descriptions, but lack constraints (enums, ranges, patterns). Output schemas are not documented, LLMs cannot predict what fields to expect. Error handling is minimal; no recovery guidance. The codebase shows defensive programming (path traversal guards in _update_file_tag), but tool definitions do not expose this to the LLM. Composition is reasonable: tools are single-purpose and chainable (e.g., brain_retrieve → brain_related → brain_classify). However, the lack of output documentation and sparse error messaging limits production readiness.
Classify a note into one of 12 brain regions, or batch-classify all unclassified notes
Correct a misclassification and log the correction for future model improvement
Show the differences between two versions of a note
Retrieve the edit history of a note
List recently modified notes, optionally filtered by brain region
List all 12 brain regions and their descriptions
Find notes related to a given note by semantic similarity
Output schemas not documented. LLMs cannot predict response structure (fields, types, pagination). brain_store, brain_retrieve, brain_classify, and others lack documented return types.
Parameter constraints missing. 'limit' parameters lack min/max bounds (e.g., brain_retrieve limit could be 1 - 1000?). 'region' is free-form string, not enum. 'version_a' and 'version_b' in brain_diff lack format guidance.
Descriptions too generic. brain_regions (54 chars), brain_history (60 chars), brain_diff (60 chars) fall below 100-char baseline for discovery/utility tools. No guidance on when to call each tool vs. alternatives.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Retrieve notes by semantic similarity search or full-text search
Restore a note to a previous version
Store a note in the brain vault with optional semantic embedding and region classification
Error handling not exposed to LLM. Code shows defensive checks (path traversal guard, file I/O error handling in _update_file_tag) but tool definitions do not document error cases or recovery paths. LLM cannot self-correct on failure.
Mutually exclusive parameters not documented. brain_classify accepts title, path, or content (at least one required) and batch flag, but descriptions do not state the exclusivity or dependency rules. LLM may pass conflicting combinations.