Loading ToolBench
Auto-capture + structured semantic recall for AI coding agents. The memory system Claude Code should have had.
claude-cortex has 10 tools with schemas and descriptions present, but quality is inconsistent. Naming follows verb_noun convention (cortex_status, cortex_search, cortex_add_trace) which is good. However, descriptions are terse (avg ~60 chars, well below the 194-char baseline), parameter descriptions are sparse or missing, and output schemas are not documented. Tools like cortex_reconnect lack clarity on side effects. No error handling guidance visible. The knowledge graph tools (cortex_kg_*) have minimal parameter documentation. Schemas are present but incomplete, many parameters lack descriptions or constraints. No evidence of pagination support for list-returning tools.
Store verbatim content into a region/cluster
Delete a trace by ID. Irreversible.
Add a fact: subject -> predicate -> object with temporal validity
Mark a fact as no longer true
Query the knowledge graph for an entity's relationships
Knowledge graph overview: entities, triples, relationship types
Chronological timeline of facts about an entity
Descriptions are terse (avg 60 chars vs 194-char baseline). cortex_reconnect ('Force cache invalidation and reconnect to the cortex store') lacks clarity on when/why to call it and what side effects occur. cortex_delete_trace ('Delete a trace by ID. Irreversible.') does not explain recovery options or confirmation flow.
Parameter descriptions are missing or minimal. cortex_kg_query accepts 'as_of' (point-in-time query) but does not document format (YYYY-MM-DD is shown in cortex_kg_add but not here). cortex_search's 'n_results' lacks min/max bounds. cortex_add_trace's 'source' parameter has no description of its purpose or format.
Output schemas are not documented. Tools return dicts (cortex_status returns {'total_traces': int, 'regions': dict}) but LLMs cannot see the structure. cortex_search returns a list but the list element schema is not visible in the code provided. This violates the requirement that 100% of A+ tools have documented return types.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2026-07-28+ | v2 |
Force cache invalidation and reconnect to the cortex store
Semantic search across the cortex. Returns traces ranked by relevance.
Cortex overview — total traces, region and cluster breakdown
No error handling guidance. Destructive tools (cortex_delete_trace, cortex_kg_invalidate) do not document recovery paths or confirmation flows. No evidence of retryable vs fatal error classification. Agents cannot determine whether to retry or ask the user.
cortex_search and cortex_status do not document pagination. cortex_status iterates in 1000-item batches internally but exposes no limit/offset parameters to the LLM. Large cortex instances could return thousands of regions/clusters, exhausting context. No documented result cap or pagination guidance.