Hello i'm Hans, currently working at DelBay Ship co as a mechanic, I dabble in models and code by night. Building auditable local-first AI infrastructure.
I recently posted about a long run I had with my agents, and at least on person seemed interested in the fact that I was able to maintain coherence over such a long session. https://t.co/RHNvAaZQZ0
You'll have to forgive me, I'm not a programmer or developer by training or profession!
I'll need to digest what you're asking and figure out how to handle it. I just started gathering this type of info and can only go by what an agent says. Being skeptical helps to a point, but production has slowed as I'm learning as I go.
I don't have the same ability as in my day job. Building Minni or Praxis isn't like diagnosing a boat and why the engine lost RPM.
Thanks for the questions! They give me something to ponder.
I recently posted about a long run I had with my agents, and at least on person seemed interested in the fact that I was able to maintain coherence over such a long session. https://t.co/RHNvAaZQZ0
I man from the contractors point of view.. they have a bar to reach before profit is even possible, insurance, on the people there doing the work, liability, etc.. so yea, when you go an 'official route' using a contractor, you'll have that. but I do agree for that small of a room its best to do it yourself, or find someone who isnt official and with insurances to do it, could be a better rate.
I’ll DM you with the flight log.
And agree, retrieval does not equal use; it is why I had Claude run an audit on the session. I tried to keep used/not-used simple: if the recall was quoted in the transcript, then I simply considered it used; I did not infer meaning beyond that.
On rev5/git-hook scars, I currently landed a use-ledger and recall-precision (did they use the information provided). The git-hook scars are in a later slice I haven't started yet; but yes, it is a very easy method to test quickly so I know that the work is going in the right direction.
Memory vs. harness: specifically on this run, it was Claude Desktop + Minni Plugin; Praxis was what the agents were building, and was not the host. And I posit that agents lean on what their host provides first, and ignore what already is provided by the host when using a plugin, so in this session it ignored Minni:threads, prepare-task, handoffs, etc., in favor of Claude Desktop’s (the host) provided tooling.
Minni harvests the session when the platform performs a compaction, then reinjects the summary package on the next prompt.
for now I use jev to create the harvest injection leveraging hooks. It uses the raw session transcript and provides the important bits. As of now there is a conflict with the agents’ platform’s method of compaction. But it’s a work in progress
I recently posted about a long session I had with coding agents, and there was small interest in further information, this is what I came up with for now.
I recently posted about a long run I had with my agents, and at least on person seemed interested in the fact that I was able to maintain coherence over such a long session. https://t.co/RHNvAaZQZ0
@morganlinton One day, as of now what I’ve got is paid off and takes me to work.
Maybe fly out somewhere with less salt and road trip back when the time comes.
I used Minni while working on another project. I’ve now another list of 'work to be done' on Minni; this session on what I'm calling Praxis was 58 hours, and 3.5mil out.
Honestly I haven't written a real pitch for it yet... mostly just dogfooding it in the trenches.
At its core, Minni is a local-first, agent-agnostic memory substrate (SQLite/FAISS over MCP) built so context survives multi-day builds and harness switches.
But here:
Minni: A local-first, auditable MCP memory substrate with active recall guards and compaction harvesting so coding agents don't lose their minds across 50+ hour builds.