2.11M views brought only 150 stars: how a DeepSeek Harness memory plugin ate three months of traffic by arriving 8 days early

dsh-memory-evolve is a Chinese long-term memory plugin for DeepSeek Harness (DSH), maintained by one person. The author has only this one repo on GitHub. It gained 7 stars in 164 days, yet the author hasn't committed a single line in those days. A first-part Bilibili tutorial used it for an install demo, and it was then pushed onto GitHub's JavaScript daily chart. What got it picked was arriving 8 days before DSH's official launch, three months earlier.

Danmaku at minute 32

On the morning of Nov 2026, 1 10:22, Bilibili UP host "Git源宝" posted a 50-minute video, "[The Most Complete] DeepSeek Harness Detailed Tutorial." As of 11-06, the video has 2.11M views, 187k favorites, and 10k+ comments, and peaked at No. 4 on the site-wide ranking.

The video starts the "plugin in practice" section at 30:50. Around minutes 32 and 33, the danmaku started flooding with the same name: "dsh-memory-evolve" and "csyangwen/dsh-memory-evolve." Some viewers shared an easier install: "You can tell Ds 'help me install dsh-memory-evolve' and it installs itself"; others got stuck: "Now it won't install." (How long the video covers it, and whether it's the main pick, can only be inferred from the danmaku timing; the subtitle text itself wasn't checked.)

The next day it gained 7 stars, and the day after 31. On 11-03 it squeezed onto GitHub Trending's JavaScript daily chart at No. 5,11-05, with +52 in a single day, its biggest day since the repo was created.

Data: GitHub day-by-day star records and Trending daily charts (11-06); trendoday's chart records.

A comparison with similar repos shows this isn't the whole category rising: over the same period (11-02 to 11-06), the biggest rival dsh-mnemon gained 78, dsh-fish-memory gained 33, dsh-meow-memory gained 22, and this one gained 144. The video also demoed the plugin marketplace dsh-market, whose daily gains in the same period rose from about 150 to about 230. The two match, so the traffic is from this video.

2.11M views brought roughly 150 stars, under one ten-thousandth. The likely reason: Bilibili viewers said "help me install it" to DSH, installed it, and left, with few going to GitHub to star it (inferred). Real install counts are unknown, since DSH plugins pull straight from GitHub with no download counter.

What it is

DSH is the agent framework DeepSeek open-sourced with V4 Pro in 2026-08-13. It now has about 270k stars and claims "everything is a plugin." Its official docs themselves say "cross-session memory is unreliable by default and needs plugin persistence." Start a new session or come back a few days later, and the AI doesn't remember you.

dsh-memory-evolve fills that gap. Install is one line:

`` dsh plugin --profile web add github:csyangwen/dsh-memory-evolve ``

Once installed, it saves your preferences, project conventions, and pitfalls in five tracks (user profile, global facts, key project memories, project log, and daily log), and conclusions tied to a project can be marked as valid only on a specific git branch. The most famous example in the README: "Note: this project's deploy port is 8080". The AI first puts it into a pending-confirmation queue, and only writes it once you nod. Most similar plugins have the AI extract and write memories itself; this one is one of the few that asks the user to confirm.

On top of that it packs in a lot more: four-track todos, self-evolving skills, pulling several AI sessions into a shared "collaboration room" to split work, dispatching heavy jobs to Kimi / Codex / Grok, cross-computer memory sync, prompt management, and session review. Most features are off by default. If the plugin stops DSH from starting, add disabled: true to the config to turn it off first. The README is in Chinese, organized by scenario (who it's for, how to use it, what you get), with an English version alongside.

Why this one: it arrived 8 days early

Its first commit is dated 2026-08-05 and reads "Hermes-style layered memory and self-evolution plugin for DSH," 8 days before DSH's public release. From 08-05 to 08-14, the author committed 20 to 39 times a day. How he got early access to DSH (beta access, the repo going from private to public, or another route) hasn't been found.

The head start paid off on the third day. On Aug 15, QbitAI published "DeepSeek Harness plugins blow up GitHub overnight: long-term memory, virtual pets, 4399 mini-games all here," and outlets such as Sina Finance reposted it. It has a dedicated section on this plugin: "For example, if you tell it today: 'This project's deploy port is 8080.' After you confirm, it writes that into project memory. A few days later, when you open a new session and ask about the deployment config, it can pick right up." That example matches the README's original line word for word, so the reporter wrote it straight from the README.

By Aug 22, in CSDN's ecosystem roundup, it had 205 stars and ranked first among DSH memory plugins (second place was dsh-mnemon with 136). After that, more than a dozen plugin directories and aggregator sites, including dsh-plugin.org, deepseek.club, dsh.plus, DeepWiki, and HotRepo, made pages for it. dsh-plugin.org's October and November comparisons of memory plugins both used it as the representative example. These sites generate content automatically, and the author didn't submit to any of them.

The author's own promotion: we found none. X search over the last 90 days returns nothing for it or the author's name, and there's no article, video or Reddit post by the author. The effort went into the product and docs, merged a few community PRs, 10-29 and adapted it for DSH 0.2.1.

So this is how the climb went: early launch → named in the first wave of press → top star count among memory plugins → became the default answer for DSH memory plugins in the Chinese community → a top tutorial picked it for the plugin demo → the language daily chart amplified it once more.

Lots of features didn't win it

Its most visible pitch is the all-in-one bundle, but the evidence says nobody came for that. The QbitAI piece only covers memory and git branches, and the comments only care about whether it installs and whether it remembers. A third-party comparison also warns that it installs from GitHub and can hit missing build artifacts.

Three plain things did the work:

• Early: the platform officially admitted memory was missing, and this was ready before the platform launched.
• Chinese: Chinese README and Chinese scenario guides, aimed at Bilibili and QbitAI readers.
• The name is the need: dsh-memory-evolve, platform prefix plus the requirement word, easy for directory sites to index and easy for users to ask their agent to install in one sentence.

The feature list had a different use: it fills a row in every comparison table and kept a group of power users (16 open issues, community PRs merged).

Who it's competing with

Memory is the most crowded category in the DSH ecosystem. An openEuler blog tally in Aug 21 already counted 59 of them. Star counts for 11-06 below:

Even it has been copied: brianwang2007xjtu/dsh-memory-evolve is a full re-upload with a single commit, an English description and 1 stars, yet dsh-plugin.org lists it separately. The 'memory plugin round two' article cites this copy and another same-name copy at piaobo123. There is also a string of same-name forks, plus projects that borrow its ideas, such as GraySilver/dsh-evolve-modes (200 stars) and Kytolly/dsh-evolve-in-git (82 stars).

If it were you

Verdict: study how it took the default slot, but don't build another DSH memory plugin. The memory slot had dozens of entries within a week of DSH's launch, and now a company-backed leader, hindsight (59k stars), sits on top. A newcomer can only move if a top tutorial picks them the way it picked this one, and you can't arrange that.

Good fit for: fast solo developers who already watch new agent platforms and new plugin systems (DSH, the Claude Code-style tools, new IDE plugin marketplaces) and can ship a working plugin in the launch week. Writing Chinese docs and knowing the Chinese community is a plus here.

Bad fit for: anyone who wants to build yet another all-in-one memory plugin for DSH and expects features alone to drive stars, or anyone who won't follow the platform's release rhythm and would rather polish slowly.

How to do it:

1. Pick a different slot. Go through the DSH official docs and issues for capabilities that the official side admits are missing or that users keep complaining about. Then check which slot on the plugin directory sites is still empty or only has a few dozen stars. The memory slot is already full.
2. Pick a different platform. When the next 'everything is a plugin' platform launches, copy this whole playbook: ship plugins that patch the official gaps in launch week, name them 'platform prefix + requirement word', and say what they do in the first sentence.
3. Do the part it didn't do well. 'Won't install' and 'no matching plugin found' in the comments point to its real problems. You could build a lighter version that installs reliably, ships its build artifacts and can be installed by an agent from one sentence, or specifically fix plugins breaking after DSH upgrades.
4. Write the README for agents to read: a one-line install command, a one-line disable command, and one example sentence (its 'deploy port 8080' was copied by QbitAI).

How to get users: target the press roundups that run in a platform's first wave (QbitAI, CSDN, the 'plugin explodes overnight' posts). This wave writes about this project even when its author does nothing. Watch the top Bilibili and YouTube tutorials for the platform, and proactively send the repo to the creators, or make the demo path short enough to explain in two minutes. Once the repo is 60 days old, wait for one outside exposure to push you onto the language daily chart. Reaching #5 on the JS daily chart needed only 30+ stars in a day.

When to move: within one to two weeks after the new platform launches. From launch to a full memory slot took only 8 days in the DSH case.

Signs you've read this wrong: ① dsh-memory-evolve keeps gaining dozens of stars a day with no new tutorial, meaning DSH memory demand is far from met and there is still room for newcomers; ② a DSH memory plugin that launched in November passes 1,000 stars within a month, meaning the slot isn't full; ③ DSH ships reliable cross-session memory built in, in which case this whole category shrinks, and 'patch the official gaps' depends on whether the official side fills them itself.

Data and sources: GitHub repo page and daily star counts (10-17 to 11-06), GitHub Trending JavaScript daily chart (11-04, 11-06); trendoday.com chart records; Bilibili 'Git源宝' video 2026-11-01 playback data and comments; QbitAI 2026-08-15 report (reprinted by Sina Finance); CSDN 2026-08-22 DSH ecosystem roundup, openEuler blog 2026-08-21; dsh-plugin.org memory plugin comparisons (10-12, 11-01); AI 2026-11-03 plugin ranking; repo README (11-06).