Brain-inspired long-term memory for AI agents — zero LLM during ingest or retrieval
Slowave provides 5 memory-management tools with comprehensive input schemas and clear descriptions. All tools are properly named with action verbs (slowave_activate, slowave_remember, slowave_recall, slowave_feedback, slowave_commit). Input schemas are fully defined with types, enums, descriptions, and validation constraints. However, output schemas are not documented in the source code provided, and descriptions, while present, are quite terse (20-60 characters), well below the baseline of 194 characters for production tools. Error handling and recovery guidance are not visible. Tool composition is sound: each tool has a single responsibility within the memory management lifecycle. The server correctly implements batch operations (slowave_remember and slowave_feedback accept array parameters for batch upsert), reducing sequential call overhead.
Prime working memory; opens implicit session
Close the task; form episodes
Append target-specific memory/procedure evidence
Semantic retrieval mid-task
Explicitly encode a durable typed claim
Output schemas not documented. LLMs cannot determine what fields to expect in responses, forcing them to guess at the structure and risking misaligned downstream tool calls.
Tool descriptions are terse (20 - 60 chars, median ~45 chars) vs. baseline 194 chars. Descriptions lack WHEN to use guidance, prerequisites, and WHAT is returned. E.g., 'Prime working memory; opens implicit session' does not explain what session management implies or how it affects downstream calls.
No visible error handling or recovery guidance. When slowave_recall fails to find results, or slowave_commit fails validation, the LLM has no actionable guidance on what to do next.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | <=2025-11-25 | v2 |
Parameter 'evidence' in slowave_recall has enum ['references', 'full'] but lacks explanation of what each value produces. The description should state: 'references: return citation IDs only (lower tokens); full: include reconstructed memory text (higher accuracy, more tokens).'
Parameter 'type' in slowave_remember uses enum with 10 values (fact, preference, decision, constraint, instruction, lesson, warning, open_question, task, artifact). Descriptions of each enum value are missing. LLMs cannot reason about which type to choose without guidance.
slowave_feedback has mutually exclusive parameters ('retrieval_id' vs 'items', 'memory_feedback' vs 'procedure_feedback') but descriptions do not state this. LLMs may pass both when only one is valid.
No pagination guidance for slowave_recall. If semantic retrieval returns many results, no limit or cursor mechanism is documented. Risk of context window exhaustion.