Loading ToolBench
MCP server for SyncMind — shared persistent memory for AI coding agents
SyncMind MCP has well-structured tool definitions with comprehensive parameter schemas and detailed descriptions. All 4 tools are explicitly registered with Zod schemas and descriptions. However, descriptions are verbose (exceeding the 50-200 char LLM-optimized range), parameter naming could be more consistent (e.g., 'memory_type' vs 'type'), and output schemas are not formally documented. Error handling is minimal, no recovery guidance or actionable error messages. The tool set is cohesive but lacks idempotency guarantees and confirmation patterns for destructive operations.
Mark a memory as actively used. Call this when you act on a memory — it boosts the memory's freshness score so it stays visible.
CRITICAL — MANDATORY EXIT GATE: You MUST call this tool before your session ends. This is how you hand off context to the next agent. Without this, everything you learned dies with your session. Fill in EVERY field that applies. Be detailed — the next agent only knows what you write here: - summary: What you did and WHY (not "fixed stuff" — explain the problem, approach, outcome) - decisions: Every choice you made and the reasoning ("Used X over Y because Z") - bugs: Bugs found with ROOT CAUSE ("Error X happened because Y, fixed by Z") - patterns: Reusable patterns ("Do X when Y because Z") - learnings: Gotchas, surprises, tribal knowledge the next agent needs - files_touched: What you created/modified/deleted - next_steps: What's left, blockers, dependencies
Read shared memories from SyncMind. Call this at the START of every session to see what other agents have learned. REMINDER: Before your session ends, you MUST call export_session to save your full context. Don't skip this.
Save a single memory to SyncMind. Use for individual insights discovered DURING work. IMPORTANT — Write useful context, not thin summaries: BAD: "Fixed CORS error" GOOD: "Fixed CORS error in /api/auth — root cause was missing OPTIONS handler. Added explicit OPTIONS export to route.ts. This affects all API routes that receive cross-origin requests." BAD: "Auth uses JWT" GOOD: "Chose JWT over sessions for auth because the app is stateless, deployed to 3 regions, and sessions would require sticky routing or a shared store." Include: WHAT happened, WHY, HOW it was resolved, and WHAT other agents should know. For end-of-session context dumps, use export_session instead — it captures everything in a structured format.
Descriptions are verbose (300-500+ chars) and exceed LLM-optimized range (50-200 chars). Baseline: avg 194 chars. This wastes tokens and buries key intent signals.
Output schemas are not formally documented. LLMs cannot plan downstream calls or extract structured data without knowing what fields to expect. read_memories returns text only; write_memory/bump_memory/export_session return no documented structure.
No error handling or recovery guidance. If /api/memories returns 404, 500, or auth error, the tool returns raw JSON with no actionable next steps for the LLM. Missing pattern: recovery-guide.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
write_memory and export_session lack confirmation or dry-run patterns. These are state-modifying operations; agents should confirm before executing. No idempotency guarantees documented.
Parameter naming inconsistency: 'type' vs 'memory_type' across tools. read_memories uses 'type', write_memory uses 'type' with default 'learning', export_session uses no type param. Inconsistent naming forces LLM reasoning overhead.