Local-first AI memory layer with hybrid vector + full-text retrieval. No API keys needed. Built-in MCP server for Claude, Cursor, and OpenClaw.
cortex-memory demonstrates solid tool design with explicit schemas, consistent naming patterns, and clear descriptions. All 5 tools are properly registered with input schemas in src/mcp-server.ts. Naming follows verb_noun convention (memory_save, memory_search, memory_forget). Descriptions range from 94-173 chars, within baseline (p10=34, p90=392). However, several parameters lack descriptions, and output schemas are completely undocumented. Error handling is minimal, no recovery guidance, no actionable error messages. The tools follow single-responsibility principle well, but composition and schema documentation gaps prevent an A-grade score.
Get a summary of all stored memories and database statistics.
Delete a memory by its ID.
Returns top 10 most-accessed memories and top 10 highest-importance memories — a "what I know well" summary.
Save a new memory. Use this to remember facts, preferences, procedures, decisions, lessons, or events for later recall.
Search memories using hybrid retrieval (semantic similarity + full-text + recency + importance). Use this to recall relevant information.
Output schemas completely undocumented. LLMs cannot predict what fields memory_search returns (e.g., does it include 'score', 'timestamp', 'tags'?). Same for memory_context stats structure and memory_reflect format. Forces agents to guess or parse unstructured text.
memory_context and memory_reflect have empty inputSchema properties {} with no documentation. Users don't know these tools exist or what they do beyond the 10-50 char description. Descriptions are too terse (50-65 chars vs baseline 194 chars average).
No error handling or recovery guidance in CallToolRequestSchema handler. If memory_save fails due to invalid namespace or importance > 1.0, returns generic exception. LLM receives no hint to validate input or retry with different params.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 53 | 2026-07-28+ | v2 |
| 2026-03-09 | D | 58 | - | v1 |
memory_save does not document output structure. Code shows it returns '{type: "text", text: `Saved memory ${memory.id}...`}' but schema never declares this. LLM cannot chain to downstream operations expecting memory.id in structured form.
memory_search result display uses string formatting '[score] (type) content [tags] [project] (id)' with no struct. LLM must parse free-form text to extract individual fields like score, type, tags. This is error-prone and wastes tokens on parsing logic.
No input validation visible in handler. If 'importance' is passed as 2.5 (violates 0.0 - 1.0 constraint), or 'namespace' is 'invalid_region', no validation error is returned with guidance. LLM receives silent failure or cryptic exception.
memory_forget does not clearly document consequences of deletion (irreversible, affects downstream references). No confirmation step or dry-run option. A destructive tool should warn the LLM explicitly.
memory_save 'metadata' parameter is typed as 'object' with optional nested properties (project, confidence, supersedes, expiresAt) but marked 'Optional metadata' (65 chars). No description of what metadata controls or when to use it. Ambiguous, LLM guesses.