Agent Memory (Knowledge Base)

Agent memory (/knowledge, right under AdWitch AI in the sidebar) is the team’s durable memory. It stores facts about the business, offers, audiences, preferences and measured results so the Agent can reuse them in later conversations.

Each entry records a source, confidence and provenance:

• You said it — supplied by a team member.
• Meta Ads data — measured in an ad account.
• Agent hypothesis — an unverified idea, not loaded into the agent’s prompt automatically.
• Measured by AdWitch — the before/after result of a change, added by the learning loop. These entries are read-only.

The page can be searched and filtered by provenance or category. Current categories include GEO, offer, creative, technical, targeting, optimisation, strategy, blocking and A/B-test results.

Knowledge Tab

Manually

1. Open /knowledge (the Knowledge tab).
2. Click Add knowledge next to the tabs.
3. Enter the title and content, then choose category, confidence and source. Subcategory, tags and evidence are optional. Measured entries cannot be created manually.
4. Save the entry.

Members with write permission can open an existing card to edit it or delete it.

Via AI Agent

The Agent uses recall_knowledge to find relevant team knowledge while working and save_knowledge to store new durable knowledge only with your consent. For example:

“Save this as a confirmed team fact: our primary market is Germany and our target CPA is €30.”

You can also ask: “What do you already know about this client’s offer and audience?”

Ad Account Scope

Every entry is either team-wide or bound to one ad account. Team-wide entries are read in every conversation and in every account’s Autopilot cycles. An account entry is read in chats pinned to that account and in its Autopilot cycles; chats pinned to other accounts do not see it. Unpinned chats may read entries from all accounts the member can access, with account labels.

• Facts the Agent saves while a client profile bound to an ad account is pinned belong to that account; say explicitly that a fact is about the business as a whole to keep it team-wide.
• Automatic entries (Autopilot outcomes, creative and A/B test results, measured learnings) are bound to the account they were measured on. Autopilot writes an outcome fact only when the result is backed by conversion evidence, and keeps a bounded number of recent facts per account and action type; entries stated by a team member are never displaced by automatic ones.
• On the Knowledge tab, bound entries show their account; the account filter narrows the list to one account or to team-wide entries, and the add/edit form has an Applies to field.
• Members whose access is limited to some accounts see team-wide entries and the entries of their accounts only.

Pending Insights and Activity

Agent hypotheses awaiting review appear in the Pending review tab (/knowledge/pending) when any exist. Approve an insight to add it to durable knowledge or reject it. The Activity tab (/knowledge/activity) shows recent knowledge events and lets permitted users remove an automatically learned insight from knowledge; measured entries remain read-only.

Client Profiles

The Profiles tab (/knowledge/profiles) stores structured client profiles for businesses or offers: brand, website, geography, targets, voice and related fields, with source and date per field. A profile only attaches to a chat through the chat’s profile pin (by choosing it, creating it or scanning a website there). Binding a profile to an account does not attach it to chats automatically.

Start in chat with: “Create a client profile for this business and scan https://example.com.” Website scanning can fill details such as brand, tagline, colours and offers; alternatively, answer the Agent’s questions manually. Agencies can keep a separate profile for each client and pin the right one to each conversation.