Memory Manager

**Quick judgment** Solid, opinionated local memory architecture for OpenClaw agents that prioritizes structure, transparency, and privacy over advanced retrieval features. Strong fit for technically comfortable users who want durable, auditable memory and are willing to work with shell scripts and markdown. **What it does well** - Implements a three-tier memory model (episodic / semantic / procedural), which is more aligned with how agents and humans reason than a flat history or single vector index. - Provides lifecycle tooling: initialization, compression-risk detection, snapshots, organization, search, stats, and manual categorization. - Stays entirely local and file-based, making data easy to inspect, back up, and version-control, with no API costs or external dependencies. - Includes a clear heartbeat pattern (detect → snapshot → organize) to help agents manage memory proactively over time. **Key limitations / risks** - v1.0 only supports basic keyword search; no embeddings, re-ranking, or knowledge-graph retrieval yet, which limits semantic recall quality compared to modern vector systems. - Organization and categorization still rely heavily on manual effort or simple heuristics; misclassification can degrade retrieval quality. - File-system operations (organize, snapshot, migration) always carry some risk of data loss or confusion if paths are misconfigured or if users customize the directory layout. - CLI- and shell-centric UX assumes a Unix-like environment and users comfortable with scripts and cron/heartbeat scheduling. **Recommended scenarios** - Long-lived or autonomous agents where memory integrity and loss-prevention are critical, and the operator values full local control of data. - Power users or developers who want a clear, inspectable memory layout they can integrate with their own tools or version control. - Early adopters willing to invest in structured memory now, with an eye toward future upgrades (auto-categorization, embeddings, graph-based retrieval) as the roadmap matures. Less ideal for: - Users who need plug-and-play, highly semantic retrieval today without touching shell scripts. - Environments where agents cannot reliably access a POSIX-like filesystem or run local scripts.