MCP server for persistent memory with JSON storage and atomic fact operations
mcp-prose-memory has 2 well-defined tools with complete input schemas and descriptions. Naming is clear and verb-based ('memory', 'memory_context'). However, output schemas are not documented, error handling guidance is minimal, and parameter descriptions lack context about validation rules and interdependencies. The 'memory' tool is well-scoped with clear enum constraints; 'memory_context' is simpler and read-only annotated. Both tools have reasonable descriptions (100-150 chars range), but neither explains what the LLM should do if a command fails or how to interpret structured responses. Parameter descriptions exist but are brief and lack guidance on format constraints or when fields are required vs optional based on the command selected.
Manage persistent memory across sessions. Commands: - view: Show all memories (optionally filter by section) - add: Add a new fact to a section - upsert: Add or replace a structured fact by key - remove: Remove a fact by section and line number - replace: Update a fact by section and line number Sections: work, personal, top_of_mind, history, instructions, user_profile, user_preferences, eyra_project, devices_environment, workflows, writing_style, long_term_tasks, do_not_forget Examples: - {"command": "view"} - {"command": "view", "section": "work"} - {"command": "add", "section": "personal", "fact": "Lives in Berlin"} - {"command": "upsert", "section": "user_preferences", "key": "answer_style", "value": "Prefers concise answers"} - {"command": "remove", "section": "work", "line": 3} - {"command": "replace", "section": "top_of_mind", "line": 1, "fact": "Working on new project"}
Load memory for session start. Supports compact bounded context for clients with small prompt budgets.
Output schemas not documented. LLMs cannot predict the structure of responses from 'memory', 'memory_context', or individual command handlers (add/upsert/remove/replace/view). This forces LLMs to make assumptions about field names, data types, and nesting, increasing hallucination risk and preventing confident downstream tool chaining.
Parameter interdependencies undocumented. The 'memory' tool's command enum determines which parameters are required: 'add' requires 'section' and 'fact'; 'upsert' requires 'section', 'key', and 'value'; 'remove'/'replace' require 'section' and 'line'. These dependencies are NOT stated in individual parameter descriptions, forcing LLMs to infer the rules and increasing incorrect invocations.
Error handling lacks recovery guidance. Code shows graceful error handling (normalize functions, validation), but tool descriptions do not tell LLMs what to do on failure. E.g., 'If the section does not exist, what should the LLM try next?' 'If a line number is out of bounds, should it retry with a different line?' Without these hints, failed calls dead-end agent plans.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 66 | 2026-07-28+ | v2 |
Parameter descriptions lack format/constraint guidance. E.g., 'fact' is maxLength 300 but the description does not mention this limit or advise on what to do if a fact is too long. 'line' is 1-indexed minimum 1, but the description does not explain the indexing convention (is it 0-indexed or 1-indexed in the UI?). LLMs frequently misinterpret undocumented constraints.
Conditional parameter optionality not explicit. 'section' is described as 'optional for view, required for add/remove/replace', but this is buried in the description and not enforced by the schema's required array or conditional logic. An LLM might pass 'view' without 'section' or 'add' without 'section' and receive unclear errors.