MCP server for portable LLM memory — store, retrieve, and manage personal facts across any LLM provider
The Memory MCP server has a clear, focused purpose with 5 well-named tools. Tool names follow verb_noun convention (remember, recall, forget, list_memories, get_changelog), which is good. However, descriptions are minimal (10-50 chars), parameter descriptions lack actionable detail and format constraints, output schemas are not documented, and error handling guidance is absent. Input schemas exist and have type definitions, but are missing key constraints (enums, min/max ranges, format patterns). No evidence of parameter validation rules or recovery guidance. This is a functional but under-specified implementation typical of many community servers.
Add a new custom category to the schema. Custom categories can be used like built-in ones. Args: - name (string): Category name (lowercase, hyphen-separated) - description (string): Human-readable description - hints (string[], optional): Extraction hints for the LLM - examples (string[], optional): Few-shot example facts Returns: The added category.
Accumulate draft facts during a conversation WITHOUT saving them to the persistent store yet. Use this throughout a conversation whenever you learn something new about the user. Drafts are held in memory and only committed when the user calls memory_save_all_facts. This is the recommended flow: 1. Call memory_create_facts throughout the conversation as you learn things 2. Call memory_review_draft_facts to show the user what you've gathered 3. Call memory_save_all_facts to commit them (or memory_discard_draft_facts to drop them) Args: - facts (array): Array of facts with content, category, optional confidence/tags - source_provider (string, optional): Which LLM provider this came from Returns: The draft facts with temporary IDs. These are NOT yet persisted.
Permanently delete a stored fact. Use with caution — this cannot be undone. Args: - fact_id (string): ID of fact to delete Returns: Confirmation of deletion.
Discard all or specific draft facts without saving them. Use this if the user decides not to commit facts you've gathered. Args: - draft_ids (array, optional): Specific draft IDs to discard (if omitted, discards all) Returns: Count of discarded drafts.
Descriptions are below 50 characters and lack WHEN to use guidance. Example: 'Store a personal fact or memory' (34 chars) does not explain when to call remember vs update, whether it deduplicates, or what happens on duplicate input.
Input parameters lack format constraints and validation rules. 'fact' accepts any string with no length/content rules. 'category' and 'tags' have no enum, pattern, or max-length guidance. 'query' in recall has no length bounds or syntax rules. 'limit' and 'minScore' in recall have no explicit min/max numeric ranges documented.
Output schemas are not documented anywhere in visible code. LLMs cannot predict what remember() returns, what structure recall() results have, or what fields get_changelog entries contain. This violates the pattern that 'every tool must document its return type'.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 11 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 29 | - | v1 |
Export all stored facts as JSON (optionally encrypted). Useful for backup, migration, or sharing the memory profile. Args: - encrypt (boolean, optional): Encrypt the export with a passphrase (default false) - passphrase (string, optional): Passphrase for encryption (required if encrypt=true) Returns: JSON export of all facts and metadata.
Retrieve the current schema (category definitions) for the memory store. Shows all available categories with descriptions, hints, and examples. Returns: Array of category objects with descriptions and hints.
Get statistics about the memory store — total facts, facts per category, pending vs confirmed, etc. Useful for status reports. Returns: Object with stats.
Import facts from a JSON export. Imported facts are merged with existing data (duplicates are detected and flagged). Args: - data (string): JSON string of facts to import (or encrypted blob) - passphrase (string, optional): Passphrase if data is encrypted - merge_strategy (string, optional): How to handle duplicates ('skip', 'replace', 'merge') Returns: Import summary with count of imported facts and any conflicts.
List all stored facts with optional filtering by category or status. Useful for reviewing the user's entire memory profile or auditing stored facts. Args: - category (string, optional): Filter to specific category - status (string, optional): Filter by status ('pending', 'confirmed', 'rejected') - limit (number, optional): Max facts to return (default 50) - offset (number, optional): Pagination offset (default 0) Returns: Array of facts matching filters, sorted by creation date (newest first).
Show all draft facts accumulated during this conversation that haven't been saved yet. Use this to let the user review what you've gathered before committing to the persistent store. Returns: All current draft facts, grouped by category.
Commit all or specific draft facts to the persistent store. This is the only way to move facts from temporary draft status into the permanent memory. Args: - draft_ids (array, optional): Specific draft IDs to save (if omitted, saves all) - status (string, optional): Status to assign ('pending', 'confirmed', 'rejected') - source_conversation (string, optional): Tag these facts with a conversation ID Returns: The saved Fact objects.
Search all stored facts by keyword, category, or tag. Returns matching facts sorted by relevance. Useful for retrieving relevant context before storing new facts, or before a conversation starts. Args: - query (string): Search query (keywords, partial fact content, tags) - category (string, optional): Filter to specific category - limit (number, optional): Max results to return (default 20) Returns: Array of matching facts sorted by relevance, with similarity scores.
Store a new personal fact about the user. Use this whenever you learn something new about the user during conversation — their preferences, background, goals, relationships, work context, etc. Facts are stored locally in the user's memory profile and persist across conversations and LLM providers. Args: - content (string): The fact to store, written as a clear statement - category (string): Category — one of: work, personal, technical, preferences, goals, health, values, social, finance, travel, education, context - confidence (number, optional): How confident you are (0.0-1.0, default 0.8) - source_provider (string, optional): Which LLM provider this came from - tags (string[], optional): Searchable tags - supersedes (string, optional): ID of a fact this replaces (corrects) Returns: The stored fact object with its assigned ID.
Store multiple personal facts at once. Use this after a rich conversation where you've learned several things about the user, rather than calling memory_store_fact repeatedly. Args: - facts (array): Array of fact objects, each with content, category, and optional confidence/tags/source_provider - source_provider (string, optional): Default provider for all facts Returns: Summary of stored facts with their IDs.
Update an existing fact's content, category, confidence, tags, or status. Use this to correct, refine, or change the status of stored facts. Args: - fact_id (string): ID of fact to update - content (string, optional): New content - category (string, optional): New category - confidence (number, optional): New confidence - status (string, optional): New status ('pending', 'confirmed', 'rejected') - tags (string[], optional): New tags Returns: The updated fact.
No error handling guidance. If forget() is called with a non-existent ID, the response should be 'Memory with ID xyz not found. Use list_memories() to discover valid IDs.' Currently no evidence of such recovery hints.
Parameter descriptions are generic and undescriptive. Example: 'tags' is described as 'Optional tags for searchability' (33 chars), does not specify array element type (string?), max array length, tag format (alphanumeric? spaces?), or max tag length.
Destructive tool 'forget' lacks confirmation/dry-run pattern. No evidence of a soft-delete, undo mechanism, or require_confirmation safety feature to prevent accidental deletion.
No documented pagination or result limits. 'recall' accepts 'limit' parameter but no guidance on max, default, or behavior when limit exceeds available results. 'list_memories' likewise lacks documented max results and default page size.
Missing natural-language identifier support. Tools accept 'id' parameter (opaque string) but no evidence of accepting 'fact_name' or descriptive references. Users say 'Delete the memory about my cat' not 'Delete memory ID abc123'.