Hippocampus

**Judgment:** A strong, opinionated memory backbone for OpenClaw agents that brings generative-agent-style persistence, scoring, and decay into a practical skill. Well-suited as the default long-term memory layer for serious agent setups, provided you’re comfortable with file/cron workflows and persistent user-data storage. **What it does well** - Implements a clear, research-inspired memory model (importance scoring + 0.99^days decay + semantic reinforcement) rather than ad-hoc notes. - Provides ready-to-use scripts for preprocessing, encoding, decay, recall, core-memory loading, and dashboards. - Organizes memory into user/self/relationship/world plus an indexed JSON schema, making it easy to inspect and debug. - Integrates directly with OpenClaw (cron jobs and `memorySearch.extraPaths`), so agents can use memories via existing RAG mechanisms. **Key risks / limitations** - **Privacy & data retention:** Stores rich, long-lived user data (preferences, emotional content, relationship details) on disk with no built-in encryption, access control, or compliance guidance. This is a concern in regulated or multi-user environments. - **Operational complexity:** Relies on shell scripts, cron, and a specific workspace layout; best suited to Unix-like environments and users comfortable managing background jobs and file-based indices. - **Scalability & robustness:** Uses JSON/JSONL files; very large histories or concurrent agents may stress performance and introduce race conditions unless carefully managed. - **LLM dependency:** Encoding quality (what’s summarized, how reinforcement is decided) depends heavily on the underlying LLM behavior and prompts; not guaranteed to be uniformly correct. **Recommended scenarios** - Long-lived personal or team agents where continuity and evolving knowledge about users, projects, and relationships is central. - Experimental "AI brain" stacks that combine multiple cognitive skills (hippocampus + amygdala + VTA) and benefit from the included dashboard. - Self-hosted or developer-operated environments where you control the filesystem and cron, and can set your own security measures around the stored memories. **Less suitable for** - Highly regulated or privacy-sensitive deployments without an additional security/privacy layer on top of the stored memories. - Simple or short-lived agents where the overhead of cron, indexing, and decay offers little benefit over lightweight context storage.