AI Assistant Memory, Fact Candidate Review & Cognitive Settings
Discover how Conversora's long-term semantic memory remembers customer preferences and stages learned facts for merchant review.
- Understand how Aura remembers customer preferences (sizing, color choices, past purchases) across conversations.
- Audit and approve AI-discovered 'Fact Candidates' in the merchant staging queue before they become permanent memory.
- Configure cognitive memory parameters including retention duration and auto-pruning rules.
- Comply with customer privacy and GDPR/local regulations with one-click customer memory deletion.
What This Functionality Does
The Assistant Memory engine gives your AI employee long-term memory, allowing it to remember returning shoppers and personalize their shopping experience across visits.
When a customer chats with a human sales associate in a luxury boutique, the associate remembers their sizing preferences, favorite colors, and past purchases ('Welcome back, Sarah! Are you looking for another linen blouse like the emerald one you purchased last month?').
Conversora brings this human touch to automated commerce: - Persistent Customer Profiles: Remembers customer sizes, preferred delivery addresses, and favorite styles across Instagram and Facebook. - Fact Candidate Staging: Rather than blindly trusting everything learned in chat, Aura identifies proposed facts (e.g. 'Customer is allergic to wool') and places them in a staging queue for merchant approval. - Contextual Personalization: Recommends matching accessories based on items the customer previously purchased or inquired about.
How It Works Behind the Scenes
The memory engine operates through an asynchronous extraction and vector retrieval pipeline:
- 1Fact Candidate Extraction: As a customer converses with Aura, an asynchronous background task examines the conversation transcript. If the customer discloses an enduring preference ('I always wear size Large in tops'), the extraction model proposes a FactCandidate record in the database.
- 2Merchant Review Queue: Proposed candidates appear in Admin > AI Employee > Memory under 'Pending Review'. The merchant can approve, edit, or reject the candidate.
- 3Vector Storage: Approved memories are vectorized and stored in the conversora-assistant-memory-v2 Cloudflare Vectorize index, scoped strictly by storeId and customerId.
- 4Memory Injection: When the customer initiates a new conversation weeks later, the memory engine retrieves their active memory profile and prepends it into Aura’s context window.
Before configuring this feature, confirm that your store meets the following requirements:
1. Reviewing & Approving Fact Candidates
To ensure memory remains accurate and free of noise, Conversora provides a 'Fact Candidate' review queue. Here, you can review statements the AI deduced during customer conversations before they are committed to permanent memory.
For example, if a customer said 'I am buying this shirt for my husband who is 6 feet tall', the AI might propose the fact: 'Customer's husband is 6ft tall, wears XL'. You can verify the accuracy with one click.
Open Memory Review Queue
From the admin console, navigate to AI Employee and select 'Memory & Learning'.
Inspect Pending Candidates
Review the list of proposed facts, including the customer name, source conversation snippet, and proposed memory.
Approve or Discard
Click 'Approve' to index the memory, or 'Discard' to permanently delete the candidate.
You can enable 'Auto-Approve Low Risk Facts' (such as standard clothing sizes) in Cognitive Settings to save time on routine approvals.
2. Customer Privacy & Memory Deletion
Respecting customer privacy is paramount. If a customer requests that their personal information or memory history be deleted, Conversora enables instant compliance.
Deleting a customer's memory purges all associated vector records from Cloudflare Vectorize and removes customer attribute records from the database immediately.
Locate Customer Record
Search for the customer in the Customers list or directly inside the Omnichannel Inbox.
Click 'Clear AI Memory'
In the customer profile card, select 'Clear AI Memory' and confirm.
Practical Business Scenarios
How leading merchants implement this functionality in daily operations:
Troubleshooting & Common Issues
Diagnose and resolve frequent failure points quickly:
Frequently Asked Questions
Is customer memory shared across different merchant stores?
Never. Memory vectors are strictly partitioned by storeId. A customer's preferences in Store A are completely invisible to Store B.