dsh-plugin-memory

In multi-session DSH workflows the real pain is not single-turn quality but context loss: every new session starts cold and you have to re-explain the project background, the decisions already made, and the lessons learned. dsh-plugin-memory takes a straightforward approach: while the model is working it records facts, decisions, lessons, and user preferences into plain Markdown files, and those files are automatically injected at the start of the next session for the same project so the model picks up exactly where it left off. Memory is split into two layers: main-file entries are injected in full within a character budget, while topic files are summarized to a one-line index with their full text loaded on demand, keeping context lean without losing information. Every write is a visible tool call; three proactivity levels tweak only the prompt wording, never the mechanics. Deletion requires concrete counter-evidence from the current turn, guarding against accidental removal of entries written by another session. An optional, off-by-default background distillation pass silently extracts memories after each turn, ideal for unattended long tasks.

Storage is plain Markdown: human-readable, hand-editable, git-friendly. Indexes are generated dynamically at injection time so they never drift out of sync. There is no external vector store or database to maintain. Best for developers who reopen sessions on the same project repeatedly and want the model to remember rather than re-explain: install the plugin, pick a proactivity level, and let the model do the rest.