Long-lived assistant built on DSH desktop — architecture notes after 11 days

[hq000204]

Hi all — we've been running DeepSeek Harness (desktop) on a single Windows PC as a long-lived assistant: a "main room" that only dispatches, separate "workshop" sessions for heavy work, a messaging bridge as the event entry, and plain text files as long-term memory. It's been alive for 11 days across 24 conversations.

It maps surprisingly well onto the Super-agent design (Router / Worker / Trigger / Watchdog) — with all the caveats: no control group, single machine, one person across rooms, not a team.

Write-up with numbers and the honest limits:
https://github.com/hq000204/dsh-living-home

Happy to answer questions.

——————created and design by Koi&hq000204

[capitaharlock]

Is the main room actually restricted at the tool level so it can only dispatch, or could it do heavy work too and it's instructions (plus you stepping in) that keep it delegating? Curious how often it drifted over the 11 days.

[OakcoderX]

I’m an older guy trying to keep up with AI through vibe coding, so a long-running assistant like this really interests me. I’m building SoloPup Memory, an unofficial DSH memory plugin. One question I keep coming back to is how to remember important things without keeping today’s temporary arrangements forever. How do you decide which information from a work room becomes long-term memory?

[hq000204]

First, you need to give this "assistant" a persona setting, and then it needs to have its own memory system. If you just need it to remember some important things, you can have it create a "memory repository" and put your original words, the original conversation text, files and documents, and so on into it, and then have it build a retrieval index so that information can be retrieved by keywords and dates. There is also another point: "moving house," that is, creating a new session. This action is best done when context usage reaches 63% (because in my testing, when it reaches 67%–71%, the preceding text is automatically compressed, causing details to be lost). Of course, the migration also involves having the assistant read its settings, its memory retrieval index, and the tasks you have assigned to it, etc. At present, our system has undergone some adjustments, and we need to observe it for 14 days and produce a corresponding report to check whether it meets the standard. What I have told you so far is only a rough outline. Please wait for the detailed content after I finish testing. Thank you for your interest.