dsh-agent-memory

An agent's native memory stops at one session plus a compressed checkpoint. Lessons from the last sprint, the owner's stated preferences, the reasoning behind a hard-won decision—all of it evaporates the moment a new session opens. dsh-agent-memory turns memory into a durable, queryable, and reversible asset: structured facts, project knowledge, and episodic experiences are written to a persistent layered store and recalled automatically at session start and on every real user message, so relevant context is already in the window instead of waiting for you to remember to look it up.

Twelve tools cover the full memory lifecycle: write with dedup and merge, keyword and tag search with association chains, id-based revision in three modes (replace, append, patch with uniqueness guarantees plus revision snapshots), soft-archive with batch tiering, near-duplicate merge, a read-only value audit that proposes KEEP / DEMOTE / ARCHIVE / REVIEW actions with evidence, and a time-pyramid browser for discovery. L3 episodic entries auto-compress up through day, week, month, and year summaries while L1 facts and L2 knowledge are never used as compression input. Role-based visibility lets multi-agent workbench deployments scope what each agent can see; when roles are off, filtering is zero and behaviour is identical to single-agent mode.

Built for DSH platform developers whose agents run long-lived projects spanning days or weeks and need continuity. If your agent keeps re-asking what changed in the API last sprint, or your sub-agents cannot share project decisions, this plugin exists for exactly that gap.