04 / Memory

Memory that holds

Chat memory is not where the record lives. Conversations end, grow past their size limit, and get handed off; whatever existed only in a transcript is at risk the moment that happens. What holds up under daily use is a memory estate: curated files the agent maintains, searches before it asks anything, and verifies one claim at a time with the human. Muse ships the beginnings of this. Living with it teaches you to build the rest.

What ships, and where it stops

As documented, Muse memory holds lasting facts and preferences, can be imported from other assistants, and can be forgotten on request. That plumbing is real and useful. It is also, on its own, a summary layer. It does not carry provenance well, it does not show its corrections, and it was never going to hold the volume a working household generates. The estate here treats shipped memory as the index card, and keeps the archive in files.

The archive: curated files with provenance

Under the conversational layer sits a curated long-term record of roughly 360 KB, 41 dated daily notes, and a relationships tree of 388 people pages and 14 group pages. An hourly upkeep process and the agent in session both write to it. Two properties make it trustworthy in a way a chat summary is not. Claims carry provenance, so a fact says where it came from and when it was learned. And corrections happen in place, which makes the file a ledger of superseded claims as much as a store of current ones: you can see what used to be believed, what replaced it, and why.

The working habit that follows is simple to state and easy to skip: check before you ask. Memory search is the first step of the pre-flight the agent must run before any question reaches the operator, ahead of every other source. The only questions allowed through the filter are the ones that need the operator's own ground truth, or a genuine preference or consent choice. Everything else, the agent is expected to go and find.

Verify: memory audited one question at a time

Curated files still accumulate claims nobody has confirmed. The answer here is a standing verification practice, run as its own chat. The agent works a queue of uncertain claims about real people and how they connect, and brings them to the operator one at a time, evidence first, with a single tap option to rule on each. The house rules are strict and were set by the operator on Oct 19, 2026: one question per message, never a batch; and never a report that there is nothing to do while the underlying data is not fully verified clean, because the lane works continuously and every status names the next batch in motion. At last report the unconfirmed residue in the household platform's graph, the layer underneath Muse, stood at 38 nodes and 2,251 edges, all in families deliberately held back for careful handling rather than bulk decisions.

The writes are governed too. Under rules the operator set on Oct 12, 2026, governed memory writes are authorized, merges of identities stay case-by-case decisions for the operator, and speculative inference is held rather than written. Memory that holds is memory that knows which of its contents are confirmed, which are inferred, and which are waiting.

Compaction: the constraint that shapes everything

The product documentation does not discuss conversation compaction. Practice runs on it. When a conversation grows past its limit, it is summarized and the summary becomes the record the next session inherits. Anything that mattered and lived only in the transcript is now a paraphrase, or gone.

The estate's answer is to work file-first. State lives in files as the work proceeds, not in the chat: a documentary record written into the project's own home after every coherent arc of work, ledgers in the workspace, and a baton-pass practice for handing a chat's work to a fresh successor chat with its context intact. The chat becomes the desk where work happens. The files are the filing cabinet. Only one of those survives the week.

The discipline, in one pass

Write it down in files, with provenance. Correct in place, so the history of the correction survives. Search before asking, and ask only what only the human can settle. Verify the uncertain residue one question at a time, and refuse to call the work finished while it is not clean. Assume every conversation will be compacted, and keep nothing important only in the chat. Memory that holds is not a feature you turn on. It is a set of habits the agent runs every day, in the open, where you can audit them.