dsh-memory-evolve: Cross-Session Long-Term Memory and Self-Evolution for DeepSeek Harness

What Does It Do?

It is a DSH plugin for DeepSeek Harness that gives the AI layered long-term memory plus self-evolution features such as skills, todos, session orchestration and external AI dispatch. It solves the problem of an assistant that forgets everything between sessions, loses track of parallel projects, and cannot offload heavy work. Maintained by piaobo123 under the MIT license and written in JavaScript, it ships with a WebUI management interface. Once enabled, every session gains a row of capability tabs, and the AI gains a set of memory, todo, skill and orchestration tools.

Core Features

• Five-track memory: user profile, global facts, project key memory, project logs and daily logs are stored in layers, progress is recorded automatically each turn, and important conclusions can be scoped to a single git branch.
• Four-track todos and local search: life, work, project and daily todos are isolated by working directory with priority and due dates, alongside filename and content search across local documents.
• Internal session team: spin up designer, frontend, backend and QA sub-sessions in one sentence and pull them into a shared room where messages are visible to all, can carry images, and status changes push notifications.
• External AI dispatch: send heavy jobs such as visual redesigns or code reviews asynchronously to Kimi, Codex, Grok or Hermes, get a task id immediately without blocking the current session, and have the summary written back into the project log.
• Memory sync: use a dedicated branch of your code repository as the memory remote so multiple machines recognize the same project automatically, with per-entry conflict choices of local, remote or both.
• Sentiment feedback logging: evaluations such as praise or frustration are recorded into daily and project logs so the AI can later analyze which task types you are unhappy with.

How to Use This Plugin?

After enabling it in DSH, open any session to see tabs for memory, skills, todos, infinite canvas, COI scheduling, session broadcast, memory sync and session review. Many capabilities are off by default to keep the AI tool list small, so turn them on in Memory Evolve settings and enable related features together to form a complete pipeline. Chat normally and the AI writes progress into the project and daily logs at the end of each turn; when something matters, say "remember this: the deployment port for this project is 8080" and it enters a confirmation queue that only takes effect once you accept it. Ask "check memory, what architecture decision did we settle on last time" to pick up where you left off, or say "create 4 sessions: designer for visuals, frontend for pages, backend for APIs, QA for acceptance, and put them in the same room" to assemble a team, then wake idle members for the next round. For multi-machine work, open project sync in the memory sync tab and push, then pull on the other machine; if your code is public but memory should stay private, point it at a separate shared memory repository. Keep the DSH process hosting the main session running, otherwise the wake-up chain stops.