dsh-memory-plugin

Every time a DeepSeek Harness agent finishes a session, all context is gone—user preferences, project background, key conclusions, none of it carries forward. dsh-memory-plugin solves exactly that: it gives agents a disk-backed long-term memory system where facts are written to a local JSON file and remain readable across restarts, new sessions, or even a fresh terminal.

The plugin ships seven ready-to-use tools—store, search, get, list, delete, clear, stats—covering the full memory lifecycle. Search blends keyword and semantic matching without any external vector service, keeping things lightweight and fully offline. Memories can be organized by category, and writing to an existing key simply upserts it to the most-recent position. Under the hood, atomic writes, per-file write locks, mtime caching, and an inverted index keep retrieval sub-linear while staying concurrency-safe; transient errors like EACCES or EBUSY trigger automatic retries, and a corrupted file falls back to a linear scan with a backup copy so nothing is silently lost.

If you are building DSH agents that need to remember across sessions, or you simply want your agent to retain the key points of a conversation the way a human would, this plugin is the lightest path: no vector database, no extra services—install it and start remembering.