Loading ToolBench
Optional MCP server for MEMENTO. Allows MCP-compatible clients (Claude Code, Codex CLI, Cursor, etc.) to use MEMENTO's memory system as MCP tools.
memento-mcp has 11 tools with complete input schemas and descriptions, but quality is uneven. Naming follows verb_noun convention (recall, remember, decide, tag, relate). Descriptions are present (avg ~120 chars) but lack actionable context, they state WHAT but rarely WHEN or WHY. Parameters have types and defaults, but many lack descriptions (e.g., memento_timeline's 'uri' param has no description). Output schemas are not documented anywhere in the code. Error handling is absent, no recovery guidance, no categorization. Security: no secrets exposed as params (good), but no audit logging visible. Composition is reasonable (11 focused tools), but chaining is unclear, responses don't document which IDs downstream tools need. Overall: functional definitions, but below production baseline for LLM-driven agents.
Register a file (plot, table, report) in the artifact registry.
Record a methodological or architectural decision with rationale.
2-hop search: fact store → keyword bridge → wiki. Use for detailed procedural knowledge.
Log a success, failure, correction, or lesson learned.
Search the fact store (SQLite FTS5, ~5ms). Returns matching facts with timestamps.
Create a typed relation between two URIs. Types: informs, supports, contradicts, extends, built-from, related_to, precedes, follows, contemporaneous.
Store a fact in SQLite fact store. Include --keywords 'wiki:slug' to bridge to wiki pages.
Output schemas not documented. Code shows tool definitions but no return type specifications. LLMs cannot plan downstream calls or extract required fields (e.g., does memento_recall return 'uri' or 'id'? Does memento_remember return a memory ID for chaining?).
Parameter descriptions missing or incomplete. 'uri' in memento_timeline, memento_tag, memento_relate has no description. 'query' in memento_wiki_search lacks guidance on format. LLMs cannot infer what values are valid.
No error handling or recovery guidance. Code shows try/except blocks but returns no actionable error messages. If a query fails or a URI is invalid, LLM receives no hint on what to do next.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | 2026-07-28+ | v2 |
Show layer statistics (fact count, wiki pages, artifacts, decisions, tag, relations).
Tag any memory URI with a semantic tag.
Walk precedes/follows chains to reconstruct chronological order.
Search wiki pages (Markdown + Git) by content. Returns matching file paths.
Tool descriptions lack WHEN/WHY context. E.g., memento_deep_recall says '2-hop search' but doesn't explain when to prefer it over memento_recall. Agents cannot disambiguate similar tools.
No tool chaining documentation. memento_remember stores facts with keywords; memento_tag and memento_relate link URIs. But responses don't document which IDs/URIs are returned, forcing agents to guess or make extra discovery calls.