csyangwen/dsh-memory-evolve

What it is

dsh-memory-evolve is a pure plugin for DeepSeek Harness (DSH) that adds capabilities the host does not ship with, without modifying DSH core and with no runtime dependencies. Its core is a five-track memory system — user profile, global facts, project key memory, project log and daily log — so an agent can recall preferences and earlier decisions after switching sessions. Project memory entries can be scoped to a specific git branch, and logs are tagged with branch markers.

On top of memory, the plugin bundles a wider workflow layer: four-track todos (life / work / project / daily, with the project track isolated by working directory), skill self-evolution and a skill manager, an infinite canvas for files and notes, prompt injection, session bookmarks and branching at any turn, context-usage warnings, a per-session advisor reviewer, in-web and IM notifications, local file and session search, and dispatch to external agents such as Kimi, Codex, Grok and Hermes. It targets developers who run many projects in parallel and want an agent that keeps context across sessions and can coordinate sub-sessions.

Most features are off by default and are enabled in the "Memory Evolve Settings" panel.

Getting started

Prerequisite: an installed DSH instance with a profile (the README uses the web profile). The plugin ships a cordis.patch.yml bundle patch, so the host registers it automatically after installation — no manual configuration.

# 1. Install into the profile (git package address; a local directory with link: also works)
dsh plugin --profile web add github:csyangwen/dsh-memory-evolve

# 2. Restart dsh web for it to take effect

Do not additionally insert the plugin into ~/.dsh/profiles/web/cordis.patch.yml; the bundle patch already registers it and a duplicate id makes the loader fail with a duplicate loader entry id error.

To change defaults, override config by id in the profile's cordis.patch.yml (top-level form, not insert):

- id: dsh-memory-evolve
  config:
    reviewEnabled: true      # enable in-turn memory review (off by default)
    reviewInterval: 10       # review every 10 user turns

If the plugin breaks startup, disable rather than uninstall it:

- id: dsh-memory-evolve
  disabled: true

Uninstalling with dsh plugin --profile web remove dsh-memory-evolve also removes the plugin's effects. If you previously inserted the plugin manually per older docs, remove those insert lines before upgrading. Detailed feature documentation and the memory-sync guide are in the repository's detailed feature docs and memory sync docs; an English README is available.

Usage constraints

• Runs only inside DSH; the install flow assumes the dsh CLI and a named profile.
• Continuous multi-agent orchestration depends on the main session's DSH process staying alive. If the main session goes idle or DSH is closed, the wake-up chain stops until you send a message.
• Memory sync stores data in a branch of your code repository by default; a separate shared memory repository can keep memory apart from public code. Conflicts only arise when the same memory entry is edited on two machines.
• Updates follow release tags, not main. After updating you must restart dsh web; the plugin stays on a detached HEAD, so do not run git pull directly. Update checks are rate-limited to at most once per 24 hours automatically.
• The advisor observes only user-visible conversation turns (not reasoning or tool-call parameters) and injects reminders as user-voice instructions, shown as collapsed [severity] lines.
• Licensed MIT.

HotRepo's take

The only recorded ranking is the 2026-11-03 JavaScript daily list, where it reached its best position of 5 with 531 stars, 61 forks and 28 stars gained that day. That single snapshot shows real traction within the DSH plugin niche but does not establish a longer trend, and no all-language daily rank was recorded. What stands out technically is scope relative to packaging: a memory core plus agent orchestration, review, canvas and notification layers in one plugin with a documented uninstall path and tag-based update verification — worth checking whether that breadth fits your workflow before enabling the optional modules.