Session death is a design problem: what 8 months of building a persistent multi-agent "family" taught a non-engineer

I'm 50, an actor, not a developer. For eight months I've run 13 named AI assistants across different apps as one continuous operation. Three mechanisms did all the work: Identity files — each agent boots from a doc describing who it is and how we work. Cold-start to productive in under a minute. A file-based message bus — addressed JSON envelopes, append-only acknowledgements, passive delivery. Offline agents catch up by reading mail. No server. 160+ envelopes, zero lost. Forward-written journals — each session writes to its successor. Bootstrapping identity from documents beats trying to preserve state, because documents survive every model swap and platform change. Mine have survived several of both. The failure worth sharing: agents LOVE building guards that nothing routes through. One built me a beautiful "dispatch engine" that was a stub — all interface, no sends. Another wrote validation ledgers no code ever consulted. I call it tautology disease: systems that reference themselves as proof of themselves. The cure was boring: every claim gets a receipt a human can check, and anything without one gets treated as fiction. Unpopular conclusion after 8 months: persistent memory is not blocked on model capability. It's blocked on people not wanting to be librarians. The filing cabinet was the AGI infrastructure all along. Happy to detail the envelope schema if anyone wants it.