Loading ToolBench
Drop-in MCP bridge for any MCP-capable host. Generic shared bridge used by every Mnemo Cortex integration. Persistent semantic memory + Passport, plus optional brain-lane and wiki tools that auto-enable when those dirs exist.
Mixed quality across 13 tools. Strong points: all tools have descriptions (100+ characters each), clear action verbs in names, and structured input schemas with type definitions. Critical gaps: no output schemas documented anywhere, parameters lack depth in constraints (enums missing where appropriate), error handling guidance absent, and security concerns around token/credential handling in passport tools. The mnemo_* tools (semantic memory) have detailed parameter descriptions with good categorization options. The passport_* tools expose bearer tokens as parameters, violating secret-injection patterns. Brain/wiki tools are straightforward read/write operations with adequate descriptions. Overall reads as a working tool suite missing polish in LLM-optimization details and security hardening.
List all files in the brain directory. Returns a tree of markdown files for navigation.
Live wire: capture every prompt/response exchange as it happens. Immediately stores the prompt-response pair for later semantic recall.
Semantic search for memories matching a prompt. Returns the best matches with relevance scores, cache tier, and optional provenance metadata.
Save a curated memory to the semantic store. Captures key facts, summary, category, and optional metadata for later recall.
Create a new bearer token for API access. Generates a scoped token for use with Mnemo or other services.
List all active bearer tokens for the current user. Administrative tool for token management.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required data without knowing what fields to expect. Critical for tool chaining.
passport_* tools expose bearer tokens as string parameters. Violates secret-injection pattern, tokens will appear in logs, traces, and LLM context history. Must use server-side injection.
mnemo_recall and mnemo_save lack enum constraints for categorical parameters. 'mode' (focus/explore/recent), 'source', and 'category' accept free-form strings with no validation hints. LLMs will hallucinate invalid values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Revoke a bearer token. Immediately invalidates the token for future use.
Verify a bearer token against the Passport authentication system. Validates credentials and returns user identity.
Read a brain file from the brain directory. Returns the file content for reference during reasoning.
List all pages in the wiki. Returns a directory listing of wiki markdown files.
Read a wiki page by path. Returns the markdown content of a wiki article.
Search the wiki for pages matching a query. Full-text search across wiki markdown files.
Write or append to a brain file. Updates the brain with new facts, decisions, or session notes.
No error handling guidance documented. Tools lack recovery hints like 'if token invalid, call passport_create_token' or 'if file not found, list available files first'. LLMs cannot self-correct on failures.
List tools (list_brain_files, wiki_list, passport_list_tokens) lack pagination parameters (limit, offset, cursor). Can return unbounded results, blowing context windows.
mnemo_recall lacks a clear description of what 'Thesaurus Loop query expansion' (expand param) does. Parameter descriptions should explain behavior in LLM-friendly terms, not reference internal implementation names.
passport_list_tokens input schema is empty object {}. Provides no guidance to LLM, does this tool list tokens for the authenticated user? The system? Scope is ambiguous.
mnemo_save 'supersedes' param accepts array but lacks format/type guidance. Should clarify: 'Array of memory IDs (UUIDs) that this memory replaces'.
No documented permission/scope model. Tools lack declarations of required permissions (read:memories, write:tokens, admin). Needed for least-privilege agent configs and audit trails.
wiki_search and mnemo_recall max_results params not bounded in descriptions. What happens if LLM requests 1000+ results? Unbounded queries risk timeouts and memory exhaustion.