Loading ToolBench
MCP server for a brain-inspired personal memory vault with recall, remember, and correct capabilities. Markdown is the source of truth; memory index for hybrid search; error correction with tiering and full audit trail.
Second Brain MCP has 9 well-named tools with clear verb-noun patterns (recall, remember, revise_memory, etc.). Descriptions are present and substantive (avg ~150 chars), explaining WHAT each tool does and WHEN to use it. Input schemas are visible and typed with Zod. However, output schemas are NOT documented, LLMs cannot predict return structure. Parameter descriptions are present but inconsistent in detail (some lack format/range guidance). Error handling is minimal; no recovery guidance or actionable error messages visible. Security is reasonable (no secrets in params, audit trail logged), but no permission gates or scope declarations. Composition is sound, tools are single-responsibility and chainable. Overall: solid foundation, but incomplete schema documentation and weak error handling prevent higher scores.
Mark a memory as needing correction, optionally replace its body, and log the correction to the audit trail. Used for error correction and fact-checking.
Soft-delete a memory by marking status=deleted and logging to audit trail. Memory is not physically removed.
Retrieve the audit trail of all corrections, revisions, and deletions. Returns the full audit log file content.
Retrieve a single memory by path or id. Returns full memory object with frontmatter and body content.
Return statistics about the vault: total memories, breakdown by type/tier/status, index signature, and vault path.
List all active memories in the vault with optional filtering by type, tier, or status. Returns array of memory metadata.
Search and retrieve memories from the vault using keyword-based hybrid search with ranking by salience, strength, and tier. Returns top-k active memories scored by relevance.
Output schemas not documented. LLMs cannot predict return structure (fields, types, nesting). Recall returns array of {m, score} but this is inferred from code, not declared in tool schema.
Error handling lacks recovery guidance. No tool returns actionable error messages (e.g., 'Memory not found at path X. Try recall() to search by keyword'). Errors are likely raw exceptions.
No permission gates or scope declarations. Tools like delete_memory and revise_memory modify state but lack permission checks or audit scope metadata. No indication of required permissions (read:vault, write:vault, admin:vault).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Create a new memory in the vault. Accepts content, type, title, tags, tier, salience, links, emotion, confidence, and needs_review. Returns memory id and path.
Update an existing memory's body content, frontmatter fields, or both. Bumps version and logs the change to audit trail.
Parameter descriptions lack format/range guidance. 'salience' is described as 'Salience score 0-1' but no validation rule stated. 'k' in recall has no min/max bounds. LLMs may pass invalid values (k=999, salience=2.5).
No dry-run or confirmation pattern for irreversible operations. delete_memory and correct_memory modify state without a preview or confirmation step. Agents cannot verify intent before execution.