Loading ToolBench
Local MCP server that gives coding agents durable, cross-session project memory backed by a Faradium memory backend. Provides tools for recording mistakes, decisions, findings, and recalling project context.
Five tools with clear, domain-specific descriptions (10 - 20 chars each) and well-defined input schemas. Naming is verb-forward and unambiguous (record_*, recall, upload_*). However, output schemas are not documented in the visible code, parameter descriptions lack depth (e.g., no format/constraint guidance), and error handling is absent. The server is functional but lacks production-grade polish: no guidance on retryability, no pagination for recall results, no validation rules stated in descriptions.
Search this project's verified memory for facts relevant to a query.
Record an architectural / design choice you made and WHY. Only durable choices a future agent should respect, not re-litigate.
Record a non-obvious research finding worth carrying to future sessions. For things you could NOT learn by reading this repo's code — an upstream bug, undocumented API behaviour, a version incompatibility. Not for where code lives or what functions do (that's greppable).
Record an approach you TRIED that did NOT work, and why it failed. Saves a future session from repeating a dead end. Only for genuine failures with a non-obvious cause — not routine trial and error.
Save an output file you produced as a downloadable artifact for the user. For deliverables that don't belong in the PR — reports, datasets, plots, exported models. `path` is resolved relative to the repo root unless absolute. Returns a status line.
Output schemas not documented. LLMs cannot infer what recall() returns (structure, fields, pagination) or what upload_artifact() yields. This forces agents to guess and risks misuse.
Parameter descriptions lack actionable constraints. 'query' and 'limit' in recall() have no guidance on format, range, or defaults. 'path' in upload_artifact() does not specify allowed extensions, size limits (25 MB cap exists in code but not documented), or resolution rules.
No error handling guidance. Tool descriptions do not explain failure modes (e.g., 'If backend is unreachable, retry after 5s') or recovery steps. Agents cannot distinguish retryable from fatal errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
recall() lacks pagination. No offset/cursor parameter or total_count in response. Large memory stores will blow context windows; agents cannot iterate through results.
upload_artifact() does not document idempotency or side effects. Agents retrying on timeout risk duplicate uploads. Description should clarify: 'Overwrites existing file at path' or 'Returns error if file exists'.