A memory management server that stores, retrieves, searches, and deduplicates memories for AI agents with semantic and lexical ranking
memd has 9 well-structured tools with complete JSON schemas and consistent naming (memory_* prefix). All tools have descriptions and input parameters are typed with enums for memory kinds. However, descriptions are terse (10-50 chars), lack WHEN/WHY guidance, and output schemas are not documented. Error handling is absent, no recovery guidance or actionable error messages. Parameter descriptions are minimal; many lack format/constraint details. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Composition is sound (single responsibility per tool), but descriptions do not meet LLM-optimization baselines (target 50-200 chars).
Correct an existing memory
Apply exact duplicate cleanup
Preview duplicate memory groups
Soft-delete a memory
Show health information for the local memory store
Summarize memories for the workspace and agent
Retrieve relevant memories for a query
Descriptions are too terse (10-50 chars) and lack WHEN/WHY context. LLMs cannot determine when to select memory_retrieve vs memory_search, or when to call memory_dedup_preview before memory_dedup_apply. Target 50-200 chars with explicit use-case guidance.
Output schemas are not documented. Callers cannot know what fields memory_store, memory_retrieve, or memory_profile return. This forces LLMs to guess field names for downstream tool calls and risks broken chaining.
No error handling or recovery guidance. If memory_store fails due to invalid kind or missing content, the LLM receives no actionable error message. No distinction between retryable vs fatal errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Search memories by keyword
Store a new memory
Destructive tools (memory_delete, memory_dedup_apply) lack confirmation or dry-run support. An agent could irreversibly delete memories without a safety gate. No tool annotations (destructiveHint) to signal risk.
Parameter descriptions are minimal. 'Workspace identifier (optional, uses default if omitted)' appears in 9 tools but does not explain what a workspace is, when to override the default, or how it affects memory isolation. Confidence score lacks range/format guidance.