dsh-memory

If you've used DeepSeek Harness, you've probably felt this: the model is powerful, but every new session feels like a fresh start—preferences, project context, and hard-won conclusions all vanish. dsh-memory addresses exactly this pain point by giving DSH a truly persistent long-term memory, stored entirely in your own local Git repository with no dependency on hosted services or cloud vector databases.

The design is restrained and thoroughly engineered. By default, memory is disabled; only when you flip the toggle in settings does the next model call receive memory guidance. Memory is recorded as Markdown with front matter, namespaced IDs, provenance, and expiry projection, while the model only sees a bounded snapshot of summary.md. Clearing memory requires a deliberate two-step confirmation, and a recovery point is preserved before deletion—so Git history can always bring it back. This makes the clear operation a recoverable day-to-day reset, not a one-way door. An optional idle-session synchronizer processes local logs in a private workspace, defaults to workspace-write, never silently installs DSH, and keeps transcripts, prompts, and credentials out of journals.

This plugin is a great fit for anyone running long-term projects or agent experiments with DSH, or anyone who simply wants the model to remember habits and context while caring deeply about privacy and auditability. If you like managing data the Git way—versioned, reversible, and inspectable—or if you're a developer looking for a safe pattern to build local persistence inside DSH, dsh-memory is well worth a look.