MarkMem provides 7 tools with complete input schemas and descriptions. Naming follows verb_noun convention (wiki_search, wiki_read, wiki_ingest, etc.), which is clear. However, descriptions are terse (10-80 chars), lacking context on WHEN to use each tool, prerequisites, or error recovery guidance. Parameters have types and descriptions, but lack constraints (enums, ranges, patterns). Output schemas are not documented, responses are JSON strings without field-level documentation. Error handling is minimal; tools return plain strings on failure rather than structured error objects with recovery guidance. No tool annotations (readOnlyHint, destructiveHint) despite clear read/write distinctions. The wiki_supersede and wiki_review tools perform state mutations but lack confirmation/dry-run patterns.

• Document output schemas for all tools. For wiki_search, specify: {hits: [{id, title, score, snippet, type, user_id}], total_count}. For wiki_read, specify: {id, title, type, user_id, claims: [{id, text, subject, valid_from, valid_until, confidence}], body}.
• Expand descriptions to 100-200 chars. Example: 'wiki_search: Search memory pages using BM25 full-text + claim ledger. Use this to find relevant facts before reading a full page. Supports point-in-time queries (as_of=YYYY-MM-DD) for temporal reasoning.'
• Add tool annotations: readOnlyHint=true for wiki_search/wiki_read/wiki_list/wiki_history; destructiveHint=true for wiki_supersede/wiki_review; idempotentHint=true for wiki_ingest (if wait=true).
• Constrain parameters: decision enum ['accept', 'reject']; source_type enum ['conversation', 'document', 'user_input']; top_k integer min=1 max=100; include_diff boolean.
• Replace string error responses with structured objects: {error: 'not_found', message: 'Page {id} not found', recovery: 'Call wiki_list() to see available pages'}. Classify errors as retryable, user_fixable, or fatal.