Customer memory and privacy

Agent Red's Persistent Customer Memory is a four-layer system that enables the AI to build context about each customer over time. Every conversation adds to the customer's profile, and future conversations benefit from that accumulated context.

This is Agent Red's primary differentiator — no competitor has confirmed implementing per-customer vector search over historical conversation transcripts.

How the layers work

Layer 1: Customer profiles (all tiers)
    ↓ Structured data: name, purchase history, preferences
Layer 2: Conversation memory (all tiers)
    ↓ Vectorized transcripts: semantic search across past conversations
Layer 3: Cross-session learning (Professional+)
    ↓ Extracted patterns: communication style, topic preferences
Layer 4: Dedicated model training (Enterprise add-on)
    ↓ Per-customer fine-tuned AI model

Each layer builds on the previous one. More layers mean more personalized responses.

---

Memory enabled

Field	memory_enabled
Type	Toggle (on/off)
Default	On
Affects	Customer memory (Layers 1–2), response generator

Controls whether the AI stores and uses customer profiles and conversation history.

When enabled:

• Layer 1: The customer's profile (name, purchase history, preferences, previous questions) is loaded at the start of each conversation and injected into the response generator's system prompt (~250 tokens).
• Layer 2: Past conversation transcripts are vectorized and searchable. When a customer asks a follow-up question or references a previous interaction, the AI can find relevant context from earlier conversations (~300 tokens).

When disabled:

• Every conversation starts from scratch. The AI treats every customer as a first-time visitor.
• No customer data is stored between conversations.
• Existing stored data is not deleted when you disable memory — it is simply not loaded. To delete stored data, use the GDPR tools in the admin console.

When to disable:

• You are required by regulation to treat every interaction independently (some financial or medical contexts).
• You want to test the AI's performance without memory influence.

---

Pattern learning

Field	pattern_learning_enabled
Type	Toggle (on/off)
Default	Off
Tier	Professional and Enterprise
Affects	Customer memory (Layer 3), response generator

Enables cross-session learning. After each conversation, the AI analyzes the transcript and extracts patterns about the customer: communication preferences (formal vs. casual, detailed vs. concise), topic interests, and buying behavior. These patterns are stored and used to personalize future conversations.

How it works:

1. After a conversation ends, a background job analyzes the transcript using GPT-4o-mini.
2. Extracted patterns are stored with a confidence score (0.0–1.0).
3. Patterns decay over time (0.05/month) — old patterns gradually lose influence so the AI adapts to changing customer preferences.
4. High-confidence patterns are injected into the response generator's prompt (~100 tokens).

When to enable:

• You have repeat customers who interact with your store multiple times.
• You want the AI to adapt its communication style to each customer over time.
• Professional or Enterprise tier.

When to leave off:

• Most of your customers are one-time buyers.
• You prefer a consistent experience for all customers regardless of history.

---

Dedicated model training

Field	dedicated_model_enabled
Type	Toggle (on/off)
Default	Off
Tier	Enterprise only
Add-on	$299/month
Affects	Customer memory (Layer 4), response generator

Enables per-customer AI model fine-tuning. When a customer has accumulated 1,000+ conversation messages, a monthly pipeline trains a custom fine-tuned model (GPT-4o-mini) specifically for that customer's communication patterns.

How it works:

1. Monthly batch: conversations are collected, PII-scrubbed, and formatted as training data.
2. Five quality gates validate the training data (hallucination, format, tone, facts, similarity).
3. If all gates pass, the model is fine-tuned and deployed with an A/B experiment (80% control, 20% treatment).
4. After 7 days, if the treatment model performs better, it is promoted to the primary model for that customer.
5. Up to 3 model versions are retained. Rollback is available from the admin console.

When to enable:

• You have high-value customers with extensive interaction history.
• You are on Enterprise tier and want the highest level of personalization.
• You have budget for the $299/month add-on.

When to leave off:

• Most stores do not need Layer 4. Layers 1–3 provide substantial personalization at a fraction of the cost.

---

Data retention

Field	data_retention_days
Type	Number (30–unlimited)
Default	Starter: 90 days, Professional: 365 days, Enterprise: unlimited
Affects	All memory layers

How long conversation data and customer memory are retained. After this period, data is moved to archival storage and eventually deleted according to the three-tier retention policy:

1. Hot storage (Cosmos DB): Active data within retention period. Immediately searchable.
2. Warm archive (Azure Blob Cool): Data older than retention period but less than 7 years. Available on request within 24 hours.
3. Cold archive (Azure Blob Archive): Data older than 7 years. Available on request within 48 hours.

How to choose:

• 30–90 days: For stores where past interactions have limited ongoing value.
• 365 days: For stores with seasonal customers who return annually.
• Unlimited: For stores that want full customer history preserved indefinitely.

---

Consent collection

Field	consent_collection_enabled
Type	Toggle (on/off)
Default	On
Affects	All memory layers

When enabled, the widget asks customers to consent to memory features before storing their data. Customers who decline consent still get AI responses, but no data is stored between conversations (effectively disabling Layers 1–4 for that customer).

How consent works:

• Consent status is per-customer: granted, denied, or not_asked.
• If consent is denied, any previously stored data for that customer is automatically deleted (GDPR Article 17 compliance).
• Layers 2–4 require explicit consent. Layer 1 basic profile data (name from the conversation) may be retained for the duration of the conversation only.

When to disable:

• Your jurisdiction does not require explicit consent for storing conversation data (check with your legal team).
• You want a frictionless experience and are comfortable with implied consent through your privacy policy.

---

© 2026 Remaker Digital, a DBA of VanDusen & Palmeter, LLC. All rights reserved.