A copilot that forgets your brand voice every morning is a search box with extra steps. BoostEcom persists facts across conversations and across channels, so the specialists stop asking what they already know.
Three scopes
| Scope | Holds | Example | |---|---|---| | Org | Truths about the company | Brand voice, target markets | | Store | Truths about one storefront | Catalogue shape, seasonality | | User | Truths about you | Preferred language, how much detail you want |
A fact is a dotted key and a value (brand.voice,
target.market.country), never a transcript. The conversation is stored
separately; memory is the distilled part.
How a fact gets written
After a turn completes, the last exchange is distilled into 0 to 5 atomic facts. Zero is a normal outcome and the common one: most turns teach the platform nothing durable.
Writes are append-only and idempotent. Every fact carries its
provenance (which message, which channel, when) and a conflicting
fact is recorded as contradicted rather than overwriting the old one.
The history stays readable, and a bad extraction can be corrected
instead of having silently rewritten the past.
Extraction happens after the response has streamed to you, so it never delays an answer.
Facts decay
Confidence in a fact falls over time, on a half-life that depends on what kind of fact it is:
| Kind | Half-life | |---|---| | Legal, allergy, accessibility | never decays | | Brand, target market, catalogue | 180 days | | Preference, tone, language | 90 days | | Season, campaign, current focus | 30 days | | Casual | 14 days |
Anything untagged defaults to 90 days.
The reasoning is the obvious one once stated: "we run a Black Friday push" is worth acting on in November and actively wrong in March, while "this product contains nuts" is never less true than it was.
Corrections
When you correct the AI, in French, English, Spanish, Italian, Portuguese or German — the correction is detected as such and recorded as the reason for the change. Correcting it in conversation is the supported way to fix a wrong fact; you do not have to find a settings screen.
Deleting
Memory follows the store and the organization. Deleting a store removes its facts and their vector embeddings; an erasure request removes the user's. See Data & privacy.
What memory is not
It is not training data. Facts live in your organization's own rows and are read back into the prompt of your own conversations: they never leave your tenant and never reach another customer's. We do not train models on your content and do not license it to anyone; the providers process it to return a response under commercial terms that exclude training on API traffic. The full statement is in the privacy policy.