Universal AI Memory Platform - Remote MCP Server + Web Management UI for persistent memory management with neural embeddings, graph memory, and tool guardrails
The Kagura Memory Cloud MCP server has significant definition quality issues. While 9 tools are registered with basic schemas and descriptions, most descriptions are generic and under 100 characters, parameter documentation is sparse, and output schemas are completely absent. Only 3 tools (remember, recall, forget) have moderately complete input parameter documentation. The remaining 6 tools (reference, explore, update, get_context_info, list_contexts, load_guardrails) have minimal or missing parameter descriptions. No tool provides output schema documentation, which is critical for chaining. Tool naming follows verb_noun conventions (remember, recall, forget, etc.), which is good, but descriptions lack the specificity needed for LLM tool selection, they do not indicate when to use one tool vs. another similar one (e.g., recall vs explore both retrieve memory information but serve different purposes). Error handling and recovery guidance are absent from all tool descriptions.
Explore the memory graph to discover connections and relationships
Delete a memory from the context
Get information about a context including memory statistics
List all available contexts for the user
Load tool guardrails (memories with tool_trigger markers) for a context
Retrieve memories from the context using semantic search
Create a reference to a memory or external resource
Missing output schema documentation for all 9 tools. LLMs cannot plan downstream tool calls or extract required fields without knowing what each tool returns.
Tool descriptions lack specificity and context. Descriptions are generic (e.g., 'Store a memory in the context', 'Retrieve memories from the context') without indicating WHEN to use each tool or how they differ from similar operations. For example, recall uses 'semantic search' but explore traverses a graph, this critical distinction is not surfaced in the recall description.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
Store a memory in the context
Update an existing memory
Sparse parameter descriptions. 'forget' has no description for 'memory_id'. 'reference' parameters (external_url, memory_id) lack context on when each is used or required. 'get_context_info' and 'list_contexts' have zero parameter descriptions. LLMs cannot infer whether parameters are optional vs required or what valid values are.
No output schema or pagination guidance. Tools returning lists (e.g., recall with 'limit' parameter, explore with graph traversal) have no documented result structure, pagination tokens, or total count fields. This prevents the LLM from knowing how to chain results or handle large result sets.
No error handling or recovery guidance in any tool description. Descriptions do not indicate which errors are retryable, which require user input, or what the LLM should do if a memory is not found or a context_id is invalid.
Destructive tool 'forget' has no confirmation or dry-run capability documented. A tool that deletes memories should support a confirmation step or require explicit user approval.
Parameter naming inconsistency. 'context_id' and 'memory_id' are IDs, but no guidance on whether these accept human-readable names or only system IDs. 'reference' tool has optional 'external_url' but no description of its relationship to 'memory_id', are they mutually exclusive or both required?