dsh-tiered-memory

The dsh ecosystem already ships over sixty memory plugins, yet every single one is flat: user preferences, environment facts, and project context all pile into the same table with no capacity ceiling and no eviction path. The memory grows, query results get noisier, and what was meant to help the model starts becoming its least reliable input. dsh-tiered-memory ports the memory discipline proven in Hermes Agent into the dsh world, replacing unbounded append-only storage with structured tiers and hard budgets.

Three tiers form the core: a user tier for stable identity facts, an env tier for runtime environment state, and a project tier for project-specific context. Each tier carries its own character budget (default cap of 400 characters per entry). When a tier is full, writes are rejected and the caller is prompted to delete or merge first, preventing the uncontrolled growth that plagues flat stores. The mem_query command performs a semantic lookup and auto-routes to the correct tier based on the phrasing of the question; it returns an empty array when nothing matches rather than fabricating an answer. tier_summary reports usage per tier and lists cold or stale candidates, while mem_delete gives the model explicit control over eviction. The plugin never deletes data on its own, keeping the curation decision in the model's hands.

This plugin is for dsh developers who run multi-session agent pipelines and have noticed that their memory layer is becoming harder, not easier, to manage. If you want a bounded, auditable, tier-aware memory discipline instead of one undifferentiated store, dsh-tiered-memory is currently the only tiered-plus-budgeted implementation available in the dsh ecosystem.