Loading ToolBench
Persistent memory for AI coding agents powered by AdoptNow.AI. Captures AI agent conversations and stores them in a knowledge graph via MCP to optimize agent context across sessions.
ContextForge has well-structured tool definitions with complete schemas and descriptions for all 7 tools. Naming follows verb_noun convention (get_, search_, save_, add_, cleanup_, forget_). Descriptions are substantive (100-250 chars) and explain WHEN to use each tool. However, parameter descriptions lack actionable constraints (e.g., 'limit' has no min/max bounds; 'hours' in cleanup_memory is unbounded). Output schemas are not documented, LLMs cannot predict response structure. Error handling guidance is absent. No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Destructive operations (forget_workspace) lack confirmation patterns in the schema itself.
Store a specific fact about an entity. Use this to record decisions, patterns, conventions, or other knowledge that should persist.
Delete knowledge from a specific time range. Use this when the user wants to clear recent memory (e.g. 'forget the last 2 hours'). Removes entities, facts, and relationships created within the specified time window.
Completely erase all knowledge for a specific workspace. Use this when the user wants a fresh start for a project, or when stored knowledge is outdated/wrong. This is irreversible.
Get all known entities, relationships, and facts for a workspace. Use this at the start of a conversation to load relevant context from past work.
Show what is stored in the knowledge graph — list of workspaces, entity counts, fact counts, and date ranges. Use this to understand what the agent remembers.
No output schemas documented. LLMs cannot predict response structure (fields, types, pagination). Breaks tool chaining and forces agents to guess what data is returned.
Numeric parameters lack bounds. 'limit' in get_workspace_context says 'max 1000' in description but no minItems/maxItems in schema. 'hours' in cleanup_memory is unbounded, LLM could pass 999999, causing unintended data loss.
Destructive operations lack confirmation pattern. forget_workspace has a 'confirm' boolean parameter, but no dry-run or explicit recovery guidance in description. Agents may accidentally erase workspaces.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Save a completed conversation. Extracts entities, relationships, and facts into the knowledge graph. Raw text is discarded after extraction. Call this when a meaningful conversation ends.
Search the knowledge graph for entities (files, functions, libraries, errors, decisions, patterns) and past conversations by semantic similarity. Use this to recall past work, decisions, and context from previous conversations across all workspaces.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools are marked with Risk levels in metadata but MCP schema lacks formal annotations. Clients cannot auto-detect which tools are safe to retry or require user confirmation.
Error handling guidance absent. No descriptions of what errors can occur, how to recover, or what the LLM should do next. E.g., what if workspace path is invalid? What if knowledge graph is corrupted?