# Business memory

Workspace memory is what stays true no matter who is writing, which channel they arrived on or which mailbox took it: what the business does, how it refers to itself, what it will not promise. It is the broadest tier, so every line in it is applied to every piece of work — which makes it the cheapest place to be right and the most expensive place to be wrong.

- **Status:** Available
- **Audience:** both
- **In the app:** #/maya-rules, #/relationships
- **Last verified:** 2026-09-10
- **Canonical:** https://connectbyjbrh.com/docs/memory/business-memory/

## What earns a place here

The test is simple and worth applying literally: would this sentence still be true on a phone call at midnight with a customer nobody has met, from a mailbox that was added yesterday? If yes, it is workspace memory. If it needs "except when…" to survive that question, it belongs at a narrower tier.

- What the business actually sells, in the words the business uses for it.
- How it names itself, and any name it must never be called.
- Standing policy that binds every channel — what is never quoted, never confirmed, never promised without a person.
- Operating facts that hold across the whole business: the working week, the languages the business genuinely operates in, the region it serves.
- How Connect should describe its own role when a customer asks who they are speaking to.

A well-kept workspace tier is short. Long is not a sign of thoroughness here; it is a sign that lines which belong on a channel, a line or a person have collected in the one place that applies to everything.

## What does not belong here

| The line | What it does at workspace level | Where it belongs |
|---|---|---|
| "Do not call before noon" (one customer asked) | Applies to every customer, on every line | Contact tier |
| "Sign off as the accounts team" | Signs every reply from every mailbox that way | Endpoint tier — the accounts mailbox |
| "Keep replies to two sentences" | Shortens written email as well as spoken replies | Channel tier — voice |
| A price list, a policy PDF, a specification | Bloats every instruction and is never cited | [Knowledge](/docs/knowledge/sources/), which is retrieved rather than always carried |

> **Careful** Supplied documents are Knowledge, not memory. Pasting the contents of a policy into workspace memory does not make answers better grounded — it makes every instruction longer, and on the phone that is paid for directly in time-to-first-word. Knowledge is retrieved for the answer that needs it; memory is carried whether it is needed or not.

## The cost of the broadest tier

Workspace memory is present on every piece of work, which is its value and its price. On a written reply the price is mostly invisible. On a live call it is measurable: instruction size is on the record as `session.instructions_chars`, and 7,500 characters produced a first token in 1.2–1.8 seconds against 2.3–3.4 seconds at 9,600. Memory reaching a call is capped at **700 characters** for exactly this reason.

That cap is not a truncation bug you should work around. It is a decision that a caller hearing a reply a second sooner is worth more than a fifth paragraph of background the model would probably not have used. If your workspace tier is too long to fit, the right response is to move lines to narrower tiers or to Knowledge, not to fight the budget.

## Keeping it honest over time

1. Read the workspace tier end to end once a quarter, from the memory viewer or the memory sheet in [Data](/docs/files-data/).
   - Result: Lines written for a situation that no longer exists become obvious when they are read together rather than one at a time.
2. Supersede anything whose truth has changed rather than editing it in place.
   - Result: The old statement stays readable, so a reply sent last month can still be explained by what was true last month.
3. Forget anything that was only ever true of one customer or one campaign, and rewrite it at the tier it belonged to.
   - Result: Behaviour narrows on the next piece of work, without a deployment and without anything else changing.

## Questions

### How much workspace memory is too much?

There is no hard limit on what you may store, but there is a hard budget on what reaches a live call — 700 characters of memory. Beyond that, length costs you latency on the phone and dilutes the lines that matter everywhere else. If the tier no longer fits on one screen, most of it is probably channel, endpoint or Knowledge material.

### Should our price list be workspace memory?

No. A price list is Knowledge: it is supplied source material an answer must be grounded in, it changes, and it should be cited rather than recited. Workspace memory is the right place for the policy around pricing — that a quotation is never confirmed without a person, for instance — not for the numbers themselves.

### Does the Owner workspace work differently?

No. The workspace tier behaves identically for the Owner and for a customer, over the same implementation. The Owner's workspace has no plan and no allowances, but memory is not metered, so there is nothing here that differs between the two audiences.

## Related

- [The four memory tiers](https://connectbyjbrh.com/docs/memory/tiers/)
- [Channel and endpoint memory](https://connectbyjbrh.com/docs/memory/channel-context/)
- [Person memory](https://connectbyjbrh.com/docs/memory/person-memory/)
- [Knowledge sources](https://connectbyjbrh.com/docs/knowledge/sources/)
- [How memory reaches a reply](https://connectbyjbrh.com/docs/memory/memory-in-channels/)
- [Structured business memory instead of a longer prompt](https://connectbyjbrh.com/research/structured-business-memory/)

## What this page is based on

- docs-source/sources/GENERAL.md §6 — the workspace tier
- docs-source/sources/PHONE.md §7 — instructions_chars and the memory budget
- `docs-source/facts.py` — MEASURED.instructions_7500_first_token_s, memory_budget_chars