mem-rag-milvus
ReviewAudited by ClawScan on May 10, 2026.
Overview
This is a coherent memory/RAG skill, but it persistently stores assistant memories in multiple places and may send memory text or search queries to an Ollama embedding endpoint without clear opt-in or deletion coverage.
Before installing, decide what kinds of information the assistant may store. Avoid storing secrets unless you have set safe database and backup paths, understand that backup JSON files may retain deleted memories, and confirm that OLLAMA_URL points only to an embedding service you trust.
Findings (3)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
Sensitive memories may remain in backup files even after a user or agent believes the memory was deleted.
The skill automatically backs up every stored memory to a JSON file, but the delete path only deletes from the active backend and does not remove backup copies.
self._backup_to_file(content, metadata, memory_id)
...
def delete_memory(self, memory_id: int) -> bool:
self.backend.delete(memory_id)Add clear retention documentation and implement deletion or cleanup for backup files, or warn users that backups must be manually managed.
Memory contents or search queries could be sent to a local or configured embedding service, which may surprise users who expect SQLite-only local storage.
The embedding function sends the provided text to the configured Ollama endpoint; store and search call this with memory content and queries, including when SQLite is the default backend.
response = requests.post(
f"{self.ollama_url}/api/embed",
json={"model": "bge-m3", "input": text},
timeout=30
)Make embedding calls explicitly opt-in, document that memory text and queries are sent to OLLAMA_URL, and avoid calling embeddings in SQLite mode unless needed.
Users must manage dependency installation themselves, and package versions are not pinned by the skill artifacts.
The skill documents manual installation of unpinned PyPI packages, while the registry metadata only requires python3 and there is no install spec.
pip install requests pymilvus
Install dependencies from trusted sources and consider adding a pinned requirements file or install spec.