Global Memory + Evolution Engine — MCP Server. A pure-local intelligence layer for Claude Code. Zero API calls, all processing on-device.
Claude Memory provides 6 tools with reasonable descriptions (avg 150 chars) and structured schemas. Naming follows verb_noun convention (memory_retrieve, memory_store, etc.). However, parameter descriptions are sparse or generic, output schemas are undocumented, and error handling lacks recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles. Schemas present but incomplete, many parameters lack type constraints or validation rules. Descriptions are adequate but not LLM-optimized for disambiguation.
Remove stale and orphan memory entries. Removes entries with: score < 10, unused > 90 days, or orphan references. Use dry_run=true to preview what would be removed without deleting.
Generate structured debate templates for multi-perspective analysis. Returns 5 role-specific prompts (Optimist, Skeptic, Analyst, Risk, Judge) pre-filled with context and related past outcomes. Work through each role in-conversation to stress-test the decision.
Process an outcome and generate evolution signals. Compares expected vs actual results, classifies failures, adjusts confidence on linked decisions, and detects recurring patterns. Conservative mode: returns suggestions but doesn't auto-apply.
Retrieve relevant memory context for the current task. Searches local repo cache first (hot.md, recent sessions), then global memory. Returns ranked results within token budget.
Dashboard of memory system health. Returns entry counts per repo, storage size, average scores, last cleanup date, and pending evolution signals.
Output schemas undocumented. No tool returns a documented schema, LLMs cannot predict response structure or plan downstream calls. memory_retrieve, memory_store, memory_debate all lack return type documentation.
Tool annotations missing. memory_cleanup is DESTRUCTIVE (removes entries), memory_store and memory_evolve are WRITE operations, but no readOnlyHint/destructiveHint/idempotentHint annotations present. Agents cannot infer safety profile.
Parameter descriptions generic or missing. 'repo' param appears in 5 tools but described identically as 'Repository name', no guidance on format, required vs optional, or how it maps to filesystem paths. 'context_type' in memory_retrieve lacks enum constraint documentation.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Store compressed intelligence after completing a task. Compresses content, assigns scores, updates local and global indexes.
Error handling lacks recovery guidance. No tool description explains what to do on failure (e.g., 'If memory_retrieve returns no results, try broadening context_type or reducing max_tokens'). Agents cannot self-correct.
Idempotency unclear. memory_store and memory_evolve modify state but no description states whether repeated calls with identical input produce duplicate entries or are idempotent. Agents cannot safely retry on ambiguous failures.