Theseus

What this is

It's an AI collaboration system I made for my own use. Raw records left by multiple agents, conclusions I want to keep, skill entries, and services like speech-to-text and OCR all sit under one set of conventions that keep everything traceable to its source. Drafts, code, and outputs stay in their own original files; only knowledge that might be useful across projects gets distilled into long-term notes.

Saving a record and getting a conclusion back are done with remember and recall. After I switch sessions, recall gives me summaries of the hits by default, and I fetch the full text when I need it. Services have one shared control entry where I can check, start, stop, and toggle them, so I don't have to change a component's code just to turn it off.

Conclusions keep their sources

Memory has a raw layer and a long-term layer, called L1 and L2. In the raw layer, events are append-only, conversation mirrors update as new material comes in, and attachments are stored separately. The long-term layer holds the user model, the system model, topic knowledge, and cross-project experience. When something happens, the raw evidence is kept first; an agent then distills it into knowledge and writes it to the long-term layer through the command line. It does not periodically reorganize all the records on its own.

When writing a long-term note, the body has to cite the raw events, and the command has to declare those citations. If citations are missing, or a declared citation never shows up in the body, the write is rejected. remember with the long-term option saves both the event and the cited note at once. The contract only checks that citations exist; whether the cited events are real and whether the conclusions are actually supported by them is not automatically checked.

One copy of each skill

The capability library is called Armory, with entries organized into 8 areas. Each capability has exactly one copy of its content. The hosts where agents run reference that copy through symlinks, and the symlink is the entry pointing to it. Capabilities owned by other repos only get their locations registered. Usage docs for services are generated by the command line and then synced out as skills.

Registering and enabling are two different things. Whether a capability is actually enabled depends on the symlinks on disk, and the manager re-reads them every time. Removing a link does not delete the content, archiving keeps it, and only an explicit delete removes the content, leaving a deletion record. Right now 369 entries are registered, 226 of them marked active. How many are actually enabled, I have not counted.

Where it stands

• The restructure is finished. Basic acceptance has passing records for speech-to-text, OCR, memory append, read, retrieval, and the isolation switch.
• That is only basic acceptance; it does not mean every part of the system has been checked item by item.
• Routing work by difficulty existed before. That routing implementation was later archived and is not in use now, and its benefit has never been verified on Theseus.
• Context rewriting is still a manually run experiment. The goal of automatically recovering from memory after the raw information is cut has not been reached yet.