asked the telegram bot about a supplement on the channel and it reacted like we had never talked. not even five messages of history, just a blank stare. the assistant had been lobotomized and it was not a model problem. at some point someone had built a custom multiplexer to serve extra profiles on the same channel, with routing rules, a profile switching plugin, and a second profile stamping its way through every message. the bot was talking to a box that had no memory of this conversation at all.
the right move was not another custom patch. the question first was what is the correct way to run one agent instance and one profile for a telegram bot, and the answer was to stop the experiment entirely. remove the multiplex flag, remove the routing, disable the switch, clear the stale runtime state that still said a second profile was being served, and restart the single gateway. the channel came back on the main profile with the full history, exactly like a normal assistant.
the interesting part of a debugging was the stale state. the config file already said the multiplexing off, but the running process was still keyed into the old identity, and fix that in the state file. one extra profile is still on disk, dormant and harmless. the lesson is also phrased as the opposite of the usual one: when the behavior looks like memory loss, check the architecture that routes the messages before you blame the model.