Chitin

**Quick judgment** Chitin is a strong, well-thought-out personality persistence layer for OpenClaw and other agent frameworks. It’s particularly valuable if you want agents to *evolve how they work* over time, not just remember facts. The feature set (typed insights, provenance-aware scoring, decay, triggers, Carapace integration) is rich and appears coherent. **Strengths** - Provides a clear separation between job description (SOUL), episodic memory, and accumulated personality/heuristics. - Local-first SQLite design with no network dependency for core operations. - Tight OpenClaw integration via hooks and PERSONALITY.md bootstrapping, with attention to token efficiency. - Provenance-aware scoring and temporal decay model that should favor durable, operator-grounded insights. - Safety-conscious defaults for promotion to Carapace (confidence thresholds, reinforcement requirements, blocking relational insights). **Key risks / limitations** - Complexity: the insight types, provenance system, trigger semantics, and Carapace/embedding options create a non-trivial learning curve for operators and tooling authors. - External dependencies for embeddings and Carapace introduce surface area for data exposure if agents are prompt-injected into sending sensitive text. - Heavy reliance on correct operator practice (e.g., judicious use of `--force`, avoiding promotion of inappropriate content, not piping sensitive data into embedding commands). - Reliability is inferred from design (SQLite, clear CLI flows) rather than from explicit testing or production-deployment evidence. **Recommended scenarios** - Long-lived OpenClaw agents that need consistent style, preferences, and behavioral patterns across many sessions. - Research and production setups where you want an auditable, versionable store of “how the agent thinks” that can be inspected, backed up, and tuned. - Teams building sophisticated agents who are willing to invest in a structured reflection and insight-contribution workflow. **Less ideal for** - Very simple or short-lived agents where the operational overhead of reflection and insight management outweighs the benefits. - Environments where any external API calls (embeddings, Carapace) are prohibited and the added complexity over a simpler memory file isn’t justified.