Memory Tiering

**Judgement:** A solid, policy-focused memory management skill that introduces a clear HOT/WARM/COLD tiering scheme and a repeatable reorganization workflow. Best suited for agents with growing, persistent memory stores and explicit compaction phases. **What it does well:** - Provides an explicit procedure for reorganizing memory (audit → redistribute → prune/summarize → verify). - Differentiates between short-term task context, stable user/system info, and long-term archives. - Encourages summarization and pointer-based handling of credentials instead of embedding raw secrets directly. - Offers simple, language-based triggers (e.g., “Run memory tiering”) and an automatic hook after `/compact`. **Key risks and limitations:** - Relies heavily on the model’s judgment to classify information into tiers; misclassification can either bloat HOT memory or bury important details in COLD. - Aggressive pruning and summarization, if applied carelessly, can remove details needed for later debugging, auditing, or regression analysis. - Assumes a particular file-based memory layout (HOT/WARM/COLD files plus dated logs); may require adaptation for other memory backends or store schemas. - No automated safeguards beyond the model’s own “verification” step that critical information has not been lost. **Recommended scenarios:** - Long-lived agents or copilots working on multi-step or multi-day projects where memory growth is a concern. - Systems with explicit `/compact` or similar maintenance commands that can hook into a structured memory-tiering routine. - Assistants that must maintain a clear separation between ephemeral conversational context, stable user preferences, and historical project records. - Environments where developers can tolerate some manual oversight of memory reorganization, especially in early adoption phases.