dsh-mnemon-gc

The third memory tier in dsh-mnemon is powered by the mnemon engine, yet the engine itself does not determine whether a memory has been superseded by newer information. Stale memories simply accumulate alongside current facts. dsh-mnemon-gc extracts that conflict-detection capability and exposes it as a standalone governance entry point focused on correctness correction: the criterion is whether a memory has been superseded, not how long it has gone unreferenced, so historically accurate memories are never flagged for cleanup.

The workflow pairs memories within the same category (immune candidates excluded) and uses an LLM sub-agent to judge whether a newer fact overrides an older one. On a hit the older memory is marked superseded, with a causal edge recording the replacement rationale, and the original entry is retained for full traceability. A three-tier classification keeps immune entries (high importance or high access count) absolutely safe, restricts superseded entries to explicit-confirmation soft deletion, and leaves everything else in watch with no deletion. Auto-patrol runs on a configurable interval (default 24 h), is strictly report-only, and never writes to the store. Batch purge uses mnemon forget for recoverable soft deletion and automatically rejects any unmarked entry. A settings card in the client GUI allows visual editing and a memory-freshness panel with search, filtering, and supervised deletion, all with instant hot-reload.

Ideal for users and teams already managing three-tier memory with dsh-mnemon who want systematic cleanup of outdated entries without manual review. The safety model is explicit: auto-patrol is read-only, purge is recoverable and limited to marked entries, and no autoPurge option exists, keeping the risk of accidental data loss minimal.