Local Codex memory manager with an official Python SDK-based MCP server.
breathing-memory provides 7 tools with complete input schemas and descriptions. Naming follows verb_noun convention (memory_*). Descriptions are present but generic (avg 95 chars, below 194-char baseline). Parameter descriptions are minimal or absent, most params lack explanatory text beyond type/enum. No output schemas documented. Error handling and recovery guidance absent. Tools are well-composed (single responsibility) but lack the LLM-optimization depth expected of A-grade servers.
Record explicit or attributable evaluation of a concrete remembered fragment.
Directly fetch remembered fragments by fragment id or anchor id.
Load active collaboration-policy fragments within a token budget.
Fetch the most recent remembered root fragments, optionally filtered by actor and reply target.
Persist one remembered fragment and any material references used in the final answer.
Retrieve remembered fragments by lexical match and rerank by search priority.
Inspect current memory state and effective parameter values.
Parameter descriptions missing or minimal. Most params (e.g., 'query', 'limit', 'verdict') lack explanatory text. LLMs cannot infer meaning from names alone, 'query' could mean full-text search, semantic search, or SQL; 'limit' could mean result count, token budget, or time limit.
No output schemas documented. Tool descriptions state what they return (e.g., 'Retrieve remembered fragments') but do not specify the structure of the response. LLMs cannot plan downstream calls or extract fields without knowing the output shape.
No error handling or recovery guidance. Tools do not document what errors can occur, how to classify them (retryable vs. fatal), or what the LLM should do next. A failed search returns nothing; the agent has no guidance on whether to retry, adjust the query, or try a different tool.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 65 | 2026-07-28+ | v2 |
memory_fetch has ambiguous parameter semantics. Both 'fragment_id' and 'anchor_id' are optional integers with no description of the difference, when to use each, or what happens if both are provided. This forces the LLM to guess.
memory_stats has empty input schema (no properties). Description is vague ('Inspect current memory state'). LLM cannot determine what parameters control the output or what diagnostics are available.