Loading ToolBench
Advanced Telegram AI Assistant with RAG (Retrieval Augmented Generation), Long-Term Memory (mem0), and MCP (Model Context Protocol) support for GitHub and Notion integration
12 tools with reasonable naming (verb_noun pattern) and descriptions, but significant gaps in schema completeness and parameter documentation. Most tools lack detailed input schema validation, parameter constraints, and output schema documentation. Error handling is minimal. Descriptions average ~100 chars but lack actionable context for LLM selection. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels assigned. Parameter descriptions are sparse or missing entirely.
Delete specific memories. Use when user asks to forget something.
Delete ALL memories about the user. Use when user explicitly asks to forget everything. Confirm before doing this.
Get recent messages from a Telegram group. Note: Telegram bots have limited history access. For questions about past discussions, use search_knowledge_base instead.
Show the user what the bot remembers about them. Use when user asks "what do you know about me?" or "what do you remember?"
List all known Telegram groups the bot is in
List all known users who have interacted with the bot
Missing output schema documentation. No tool documents what fields/structure it returns, forcing LLMs to guess downstream field names and types. This breaks tool chaining and wastes tokens on parsing unstructured responses.
Destructive tools (forget_about, forget_everything) lack confirmation/dry-run pattern. No mechanism to prevent accidental data loss. LLMs should be required to confirm before executing irreversible operations.
Parameter descriptions are minimal or absent. 'schedule' param in schedule_recurring_message lacks format specification (e.g., cron syntax vs natural language). 'time' param in set_reminder lacks examples of valid formats. LLMs cannot infer constraints from names alone.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 46 | 2026-07-28+ | v2 |
Explicitly store something the user wants you to remember. Use when user says "remember that..." or "please remember..."
Schedule a one-time message to be sent later
Schedule a recurring message (daily, weekly, etc.)
PRIORITY TOOL: Search through indexed message history using semantic search. Use this for ANY question about past discussions, topics, decisions, or finding what was said.
Send a message to a Telegram user or group immediately
Set a reminder for the user
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. Tools marked WRITE/DESTRUCTIVE should declare this in schema for agent safety and audit trails.
Error handling is absent. No guidance on what to do if a message send fails, a channel is not found, or a reminder time is invalid. LLMs receive no recovery hints.