Hermes Agent memory as an Obsidian vault the customer owns

Notes alone took three models from 5 or fewer right to 33 to 39 of 40 on Fieldstone IT; the quote action closed the rest

Hermes Agent is an open-source agent from Nous Research: it runs a model, uses tools and keeps its own memory. Staffbox is not affiliated with Nous Research; we build on Hermes Agent and do not speak for the project. What we add is the company's own knowledge, and the test below shows how much it matters. Each group is the same model asked the same 40 test requests: without the brain, with the brain only, and (for the local models) with the brain plus the quote action.

Read this carefully. Both companies in these tests are fictional demo companies, the requests are test requests, and no customer has run Staffbox yet, so there is no real customer vault to report on. The brain-only rows isolate the vault: it took Qwen3 8B from 1 to 33 and the 27B from 2 to 39 on Fieldstone IT (Dell test box, 30 Sep). The quote action, which does the prices in code, closed the rest. The cloud model ran with notes only, so its row compares local with the quote action against cloud with notes only; it is not a ranking.

Read these scores with four caveats. Notes + calculator scored 31/40, below notes alone (33/40): a general calculator didn't help; the quote action (prices in code) did. "Unseen" = 15 distinct requests re-run 6 times (90 answers). Fieldstone's 40 standard tests are the set the quote action was built on. All Mac mini scores are on a 16 GB unit; customer pilots may run on 32 GB (the install guide: "32 GB memory recommended; 16 GB runs the 8B model"), not yet scored on 32 GB. Browse every graded answer.

Plain Markdown keeps the brain readable, portable and fixable without a developer

The brain is a Markdown/Obsidian vault the customer owns: a folder of plain text files that Staffbox builds from the client's own files at the start. Three practical reasons drive the format, all about the people who will keep it up to date.

• Anyone can read it. Markdown opens in any text editor. Obsidian is a third-party note app that opens a folder of Markdown files as a linked vault, so an office manager can browse it without a developer.
• It is not locked in. If you stop using Staffbox, the files are still files. Nothing needs an export step.
• Fixes are visible. When a draft is wrong, the cause is usually one missing or outdated line in a file. A person can find that line and change it.

The choice of format is a design decision by our engineers (our view). How Hermes Agent's own memory and agent-written skills interact with the vault on a customer unit is not yet in our public record; see what we leave switched off.

The vault holds what a new hire needs on day one, including what never to price

This is our guidance for what to gather for the quote desk, not a fixed file layout (our view).

• The current price list, with the date it was issued. Every priced line in a draft cites the price-list row and that list's date.
• The names customers actually use for products, including nicknames and old part numbers.
• Terms that change a quote: minimum orders, delivery, installation, validity period.
• Who handles which kind of request, so triage can route to a queue or a person.
• Things the worker must never price. A refusal counts as a miss on the scorecard, but a wrong price costs more than a miss.

The Staffbox repository includes a demo vault for the fictional Fieldstone IT with 40 tests (16 quote, 10 lookup, 8 triage, 6 unknown). It is a good model of what a first vault looks like, and anyone can read it. How large a vault can grow before answers slow down or degrade is not measured yet.

A wrong draft is fixed in the vault first, then kept as a test

Staffbox's sales motion is "test on past requests, scorecard, fix ladder, live only on passing workflows, monthly re-test; corrections become tests." The fix ladder goes brain first, then tool, then model.

Here is how that played out on the fictional Fieldstone IT. A customer asks for a quote on a set of items that includes one product Fieldstone does not stock. The brain tells the worker which products exist and what they cost; the quote action prices the stocked lines, cites the price-list row and date for each, and flags the unstocked item rather than swapping in the nearest SKU. That behaviour came from a measured failure. In the 5 Oct run on a generated regression set (360 requests written by a model, answers checked by code), all 6 wrong totals involved a request that also asked about an unstocked item. Since the 6 Oct fix, the quote action flags the unstocked item instead of substituting (by design; 0 wrong totals in 475 test answers). On a new trade it had never seen, an electrical distributor (14 Oct, Dell test box), the default Qwen3 8B setup scored 30 of 40 and 2 of 15 on messy requests, with 8 wrong prices in 55, because per-foot units, price levels and trade shorthand are not handled yet; that is why we score your own list first (the electrical result).

On a live unit, a person who finds a wrong draft tells us or their IT provider. If the cause is knowledge, the vault line is corrected. The failing request is then added as a test (the CLI has an addtest command) so the same mistake is caught after every later update. See monthly updates and the test set.

The vault stays the customer's, during the service and after they leave

• The brain is always the customer's.
• After the first paid month, the box and the brain are the customer's to keep.
• No one trains a model on customer data.
• On exit, the unit is returned or wiped in front of the customer, and Staffbox deletes its copies within 14 days and confirms in writing.
• For the free 30-quote score, the files are deleted within 14 days of sending your results (per our draft data policy).

Sources

• Hermes Agent repository (Nous Research, MIT), seen 13 Oct 2026.
• Hermes Agent documentation, seen 13 Oct 2026.
• Staffbox: scores from evals/results in github.com/Staffbox-ai/staffbox (MIT), 30 Sep, 2 Oct, 5 Oct and 6 Oct 2026; security sheet and the quote-desk offer (2 Oct 2026).