Standalone HTTP MCP memory server with evidence-gated recall, lifecycle, and local control portal.
Dense-Mem provides 6 tools with schemas and descriptions visible in test files. Naming follows verb_noun convention consistently (list_, get_, resolve_, eval_run_, eval_list_, eval_run_). Descriptions are present but vary in depth (44-195 chars). Input schemas are well-structured with types, required fields, and enums where appropriate. However, parameter descriptions are sparse or missing in critical tools (list_dreams has no description for 'status' enum), and output schemas are not documented. The server targets a specialized domain (hypothesis memory management) with sophisticated evaluation features, but lacks guidance on error recovery and chainability between tools. Tools are functional but not optimized for LLM decision-making at the description level.
Page through authenticated-team knowledge references for evaluation. Content is included unless metadata_only=true.
Run an isolated manual dream phase for a team during evaluation so expected dream hypotheses can be exported and scored.
Run one recall/context evaluation case through the current Dense-Mem logic and return ranked refs plus context refs.
Get a specific hypothesis dream by ID.
List hypothesis dreams with optional status filtering.
Resolve feedback on a hypothesis dream with decision and optional evidence/relationships.
list_dreams: 'status' parameter has an empty enum with no description of valid values. LLM cannot determine which statuses are supported.
No output schemas documented for any tool. LLMs cannot predict the structure of responses, forcing them to infer fields and risking downstream tool chaining failures.
eval_run_dream_cycle: 'seed_dreams' parameter has deeply nested schema (objects with sub-objects) but descriptions for nested fields are missing. LLMs will struggle to structure valid payloads.
No error handling guidance visible in tool descriptions. Tools do not document what errors can occur, whether they are retryable, or what recovery steps are available.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | B | 71 | 2025-06-18+ | v2 |
list_dreams and eval_list_knowledge_refs support pagination (cursor, limit) but descriptions do not explain cursor semantics, what 'total_count' is returned, or how to handle end-of-list.
eval_run_dream_cycle: no documentation of expected response format, success/failure semantics per item, or limits on how many hypotheses can be exported.
Destructive write tools (resolve_dream_feedback, eval_run_dream_cycle) lack idempotency guidance. No mention of whether repeated calls with same input are safe or if they risk duplicate state mutations.