Loading ToolBench
clude-mcp has 16 tools with reasonably complete schemas and descriptions. Most tools follow verb_noun naming and include input parameter documentation. However, there are critical gaps: no output schemas are documented anywhere in the codebase, parameter descriptions lack detail on formats/constraints, and error handling guidance is minimal. The tool set is well-composed (each tool has a single responsibility), but LLM usability is hampered by missing return type documentation and sparse parameter validation hints.
run type-specific decay pass (self-hosted only)
consolidation → reflection → emergence cycle
serialize memories to a portable signed bundle
format Memory[] into an LLM-ready string, local
memories created/accessed within N hours
all self_model memories
aggregate counts + average importance/decay
batch-fetch full content by ID array
NO OUTPUT SCHEMAS DOCUMENTED. All 16 tools lack any visible return type documentation in the codebase. LLMs cannot infer what fields to expect, forcing them to guess at response structure and plan downstream calls blindly. This is a critical blocker for agent reasoning.
PARAMETER DESCRIPTIONS LACK DETAIL. Input schemas are present but descriptions do not specify expected formats, ranges, character limits, or constraints. E.g., 'importance' param in store_memory is '0-1' but LLMs often pass invalid values. Descriptions should state 'importance score (numeric, range 0.0 - 1.0)' or include an enum.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 49 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 32 | - | v1 |
load a Memory Pack into the store
regex concept extraction, local, zero latency
ingest conversation sessions and convert to memories
create a typed directed edge between memories
full hybrid search → Memory objects
lightweight hybrid search → summaries only
LLM-scored importance for a description string
write a new memory with optional auto-scoring
NO ERROR HANDLING GUIDANCE. Tool descriptions do not explain failure modes or recovery actions. When recall_memories returns no results, should the LLM retry with different keywords, ask the user, or try a different tool? Lack of recovery guidance wastes agent steps.
VAGUE TOOL PURPOSE FOR COGNITION TOOLS. 'dream' description ('consolidation → reflection → emergence cycle') is esoteric and does not explain what LLM behavior should trigger it. 'decay_memories' is marked 'self-hosted only' but the description does not say what it does or when to use it.
NO PAGINATION GUIDANCE FOR LARGE RESULT SETS. 'recall_memories' and 'recall_summaries' accept a 'limit' param but no description of typical result sizes, max limit enforced, or whether pagination is supported. If agents hit result caps, they need guidance on iterating.
AMBIGUOUS PARAMETER NAMES. 'pack_data' in import_pack is described as 'base64-encoded memory pack' but lacks encoding format constraints or expected structure. 'memory_ids' arrays lack description of ID format (numeric, UUID, string).