dsh-memories

Starting every new session by re-explaining project conventions, user preferences, and where you left off is repetitive, error-prone, and breaks flow. dsh-memories borrows the memory-pipeline idea from Codex and maintains two living Markdown ledgers per workspace: MEMORY.md captures stable facts (preferences, project norms, environment details, lessons learned) while PROGRESS.md tracks completed, in-progress, and upcoming work. Both are auto-injected into the system prompt of every new session, so context is ready the moment a conversation starts.

The plugin works silently in the background: throttled to once every 30 minutes, it reads session texts from the last 14 days in the same workspace, lets an LLM distill them into four fact categories and three progress states, then consolidates into clean, deduplicated, sectioned ledgers with automatic .bak backups. Secrets, code dumps, and one-off task details are strictly excluded; entries unconfirmed for 30 days decay unless marked as pinned. On the agent side, remember and update_progress tools let the model persist knowledge mid-conversation; on the human side, /remember, /progress, and /memories slash commands give direct visibility and control.

Designed for developers running long-lived projects on DeepSeek Harness-multi-person collaboration, multi-day iterations, or anyone who is tired of copy-pasting context into every new chat. All data lives in plain Markdown, free to open, edit, and diff.