An agent suggested changing how my other agents would remember our work. I hadn't agreed. I said we should test the simpler version first.
In an early test, the memory-taker carried the suggestion forward anyway.
I spend hours working through decisions with agents. We propose things, reject them, revise them, and sometimes leave a decision open. Eventually a long conversation gets shortened so the agent can keep working. A note that preserves the wrong version of events can steer tomorrow's work.
This is the problem we're working on inside Memory Loom, the part of NLC that carries useful context forward. The new piece is called Segment Ledger.
In the version we're testing, a small, fresh agent reads a completed stretch of conversation on the user's device. The local copy contains only user and agent messages in order, without tool logs or attachments. From that, the scribe records what happened and what was decided, citing the message IDs behind each note. Code checks those citations against the conversation. Memory Loom keeps the resulting note and its source reference; the full conversation stays on the device.
We tested the revised scribe on a conversation containing accepted and rejected proposals, without giving it the answers. It kept the agent's proposal separate from my response and cited the right messages. This morning, a disposable end-to-end test also saved a note in Memory Loom that we could reopen with its supporting messages.
That is the standard I'm trying to build toward. When I come back to the work, the next agent should know where the conversation actually ended.
FELIX