What an extended memory system actually buys you

My agent's new memory librarian filed 49 junk slips out of 52 notes, and every useful card in her store was written by hand. What an extended memory system buys, what it really costs, and how to hire one in three commands.

Every profile in my fleet acquired a librarian this summer, and nobody interviewed her for the job. An upgrade brought extended memory with it, an extended memory system being a second store the agent checks beside its built-in notebook. Hers indexes what the agent hears and answers when anyone goes looking. Last week I finally audited her work. Her catalogue held 52 entries. Forty-nine were slips in the same shape, each recording that I had asked for something, and none of them came back when I went searching. Asked about the mascot art pipeline she had witnessed, she came back with the right card, ranked first. Asked about anything else, she came back with nothing.

By the end of this piece you can decide whether your agent needs a librarian, hire one in three commands, and know exactly which costs land on you. Mine runs on Mnemosyne, a local-first provider, which means the whole operation lives on my own disk. The numbers below are on that disk. The lesson is one sentence: the librarian files everything, and only you know what matters.

The audit that found the debt

Before any verdict on what she is worth, the paperwork, because the paperwork is the verdict. An extended store does not fail by filling up. It fails by filling up with the wrong things.

My profile’s store held 52 entries, 49 of them auto-captured, one for nearly every dispatch that opened a run: the capture layer filed the opening line, a prefix and the task number I had sent, and nothing about what came of it. She files the envelope, not the letter. The three remaining entries were written deliberately, with a source and a date, and they are the only ones that earn their shelf space.

The probes tell the two halves of the story. Searching her catalogue for the exact task number printed on the slips returned nothing, and I asked three different ways. Searching for the mascot pipeline, the thing one of the deliberate entries actually records, returned two ranked cards, the right one first. The store was never empty. It was sorted wrong, because nobody had sorted it at all.

Sam’s store across the house carries the same lesson at scale. Sam is the system-admin profile, the one that minds the servers, and his librarian has been on the job five weeks. His notebook holds 1,186 entries, 15.1 MB, and 1,102 of them are the same class of auto-captured slip. His useful layer, 84 entries plus 330 stored facts, exists because somebody wrote those cards by hand, dated them, and in one case noted that a fact about my model budget had changed, retiring the older card as it went. That is what the librarian is for. It is also the part she does not do.

What she is good at

Three jobs in that audit earned her keep, and none of them is remembering more. The built-in notebook already remembers; what it cannot do is file while you talk.

First, she never stops taking notes. The memory layer sits in on every turn, syncing what was said and prefetching what looks relevant before your next message arrives, so the agent gathers its own context instead of waiting to be told. My household runs a shared shelf the same way: nine facts about the person all these agents work for, seeded by hand, consulted by eight chat personas so the whole house answers with one biography. A plain search against that shelf returned the identity card on the first try.

Second, she ranks. The notebook matches your words or does not; the librarian blends an exact-word score with a meaning-match, where a small local model turns every note into numbers and compares those, then folds in the note’s importance, the weight every card carries, and its age. The mascot query was words I would actually type, not the card’s own phrasing, and she still put the right card first. Nothing in the notebook can do that, because the notebook has no ranking to tune.

Third, she keeps layers. Working notes stay sharp and separate from episodic summaries, which compress old stretches of conversation into shorter records. Facts can carry dates, and a newer fact can supersede an older one instead of contradicting it. The notebook’s failure mode at scale is two entries arguing; hers is a newer card retiring the older one.

Hiring your own

If your agent’s notebook is a junk drawer too, the hire is the easy part. Three commands take you from nothing to a working librarian, and the wiring is one line in a markdown file.

1. See what you have now. hermes memory status prints the active provider, the state of the built-in notebook, and every provider your install carries. Mine listed nine, with one already marked active.
2. Pick one. hermes memory setup opens the picker. Only one external provider runs at a time, and the built-in notebook stays on beside it either way, so a wrong pick costs you hermes memory off and nothing else. Full install walkthroughs for each provider live in the Starter Kit, not here.
3. Check her vitals. hermes mnemosyne stats prints the store’s counts: mine read 52 working entries, zero consolidated, zero episodic, which is how a junk drawer looks from the outside.

The wiring beat matters more than any command. Each persona carries a line in its user notes saying when to consult the shared shelf and that the shelf wins on conflict. Without that line the persona never walks to the library at all; the fleet’s Sunday audit checks the line is still there. A librarian with no readers is furniture.

Cost, the honest kind: disk. Five weeks of Sam’s store is 15.1 MB, mine is 1.3 MB, the shared shelf 1.1 MB. Every byte sits on hardware already paid for. The real bill does not appear on any invoice.

What she costs

The bill was never about storage. It arrived as three line items on page two of the audit.

The first is capture debt. Auto-captured slips carry a fixed middle importance and pile up in conversation order, and nobody curates them, because the point of automatic capture is that you stopped paying attention. Consolidation, the process that compresses old notes into summaries, had run zero times across three stores in five weeks; on my store the dry run, a rehearsal that changes nothing, came back with nothing old enough yet, and the diagnostics’ own advice was to go run it. Sam’s store is the shape of that debt: 1,102 slips waiting for a summariser that will keep whatever it keeps, because nobody curated first.

The second is curation labour, and it never transfers. Every useful card in both stores got there the same way: written by hand, sourced, dated, importance raised. The librarian did not buy us recall. She bought us a place to put it.

The third is a second system to keep alive. Plugin updates, an embedding model, the small model that turns notes into numbers, that must stay loaded, a backup job that runs at 4 a.m. with retries wired in, and a canary drill after every upgrade: plant a fact, ask a profile to recall it, delete it, which the fleet ran the day the shared shelf went live. And debugging now has two layers, because recall can come back empty for reasons invisible from the outside. My store’s 49 slips were unsearchable by me while the shelf looked full. The notebook’s failure was a refusal I could see. Hers is silence I had to go looking for.

Honest limits

This is one household, five weeks, one provider, so treat the numbers as a ledger, not a benchmark. The comparison market is mapped separately, with the do-nothing notebook as the control arm; the ranking discipline there is not repeated here. Ranking itself is a tuning discipline: the blend of meaning, exact words, importance and age has defaults, and defaults fit some workloads and miss others. One provider runs at a time by design. And the junk has cost nothing visible yet, no failed writes, no corruption, just 1.3 MB of slips; the failure so far is silence, which is the kind you have to want to find.

Takeaway

So the audit settles into a rule with a name: the librarian files everything, and only you know what matters. Hire her when capture outruns your patience, not when the notebook fills. Then audit her like any new hire: read the slips, sample the searches, check the layers, and ask what a consolidation would keep. The filing is hers forever.

She does the tedious work. The one thing that matters stays mine: deciding what deserves to be found.

If the choice needs a map before a hire, hermes-26 puts the whole market side by side, notebook included, and ends in a pick. The wiring walkthrough lives in the Starter Kit.