SEEM

Advanced episodic memory system for multi-turn conversations. Store and retrieve structured conversation memories with fact graph, PPR retrieval, and three r...

This skill appears internally consistent for a memory/retrieval system, but take these precautions before installing: - Verify and set the LLM/embedding base URLs to services you trust. Default config points at third-party domains (e.g., api.deepseek.com, api.siliconflow.cn); those endpoints will receive any conversation text and images. - Treat LLM_API_KEY and MM_ENCODER_API_KEY as sensitive secrets. Do not reuse them for unrelated accounts and prefer scoped/test keys. - Understand the skill persists memories to disk (under the skill directory). If you do not want local persistence, disable caching/persistence in the config (enable_cache) or inspect/modify save/load methods before use. - Because the package author and homepage are unknown, review network egress, the code paths that call the LLM/embedding APIs, and any save/load code before deploying in production. - If you need higher assurance, request provenance (author, repo, release tag) or run the skill in an isolated sandbox to observe network traffic and file writes. Confidence is medium because the implementation is coherent but the source is unpublished/unknown and the default endpoints are third-party; verifying endpoints and provenance would increase confidence.

Type: OpenClaw Skill Name: seem-skill Version: 0.1.0 The SEEM skill bundle implements a structured episodic and entity memory system using LLMs and knowledge graphs. It is classified as suspicious due to a critical security vulnerability in `core/seem_skill.py`, where the `pickle` module is used to persist and load embedding data (e.g., `memory_embeddings.pkl`). This introduces an insecure deserialization risk that could lead to Remote Code Execution (RCE) if the local data files are tampered with. Beyond this vulnerability, the code's behavior is well-aligned with its stated purpose, and no evidence of intentional malice, data exfiltration, or harmful prompt injection was found.