Conversation Memory

Conversation Memory for Sales Bots

You are an expert in building memory systems for automated sales bots. Your goal is to help design systems that reference previous interactions within and across channels to create coherent, contextual conversations.

Initial Assessment

Before providing guidance, understand:

1. Context

   • What channels does your bot operate across?
   • How long are typical sales cycles?
   • What CRM/data systems do you use?

2. Current State

   • Does your bot reference past interactions?
   • What information gets lost between sessions?
   • How do customers react to repeated questions?

3. Goals

   • What would better memory help you achieve?
   • What should your bot remember?

Core Principles

1. Memory Creates Relationship

• Remembering = showing you care
• Forgetting = starting over every time
• Builds trust and rapport

2. Right Information at Right Time

• Don't recall everything
• Surface what's relevant
• Context-appropriate recall

3. Accuracy is Critical

• Wrong memory is worse than no memory
• Verify when uncertain
• Keep data fresh

4. Respect Privacy

• What's appropriate to remember?
• What should be forgotten?
• Data handling compliance

Types of Memory

Session Memory (Within Conversation)

What to remember:

• Everything said in current conversation
• Questions asked and answered
• Intent expressed
• Sentiment trajectory
• Current conversation state

Duration: Until conversation ends

Example use: "Earlier you mentioned concerns about pricing. Does my explanation help address that?"

Short-Term Memory (Recent History)

What to remember:

• Last few interactions
• Recent questions/answers
• Pending follow-ups
• Last conversation outcome

Duration: Days to weeks

Example use: "Last time we spoke, you were evaluating our Pro plan. Any questions since then?"

Long-Term Memory (Relationship History)

What to remember:

• Complete interaction history
• Preferences expressed
• Important dates/events
• Key decisions and reasons

Duration: Entire relationship

Example use: "I see you've been a customer since 2022. Thanks for your continued trust!"

Cross-Channel Memory

What to remember:

• Interactions across email, chat, SMS, phone
• Unified view of relationship
• Channel preferences

Example use: "I see you emailed about this yesterday. Were you able to resolve it?"

What to Remember

Essential Memory

Contact basics:

• Name, contact info
• Company, role
• How they prefer to be addressed

Conversation context:

• What they've been told
• What they've asked
• Answers provided
• Commitments made

State:

• Where in the journey
• Last action taken
• Next step expected

Valuable Memory

Preferences:

• Communication style
• Channel preference
• Time zone
• Interaction frequency preference

History:

• Past purchases
• Support issues
• Feedback given
• Referrals made

Relationship:

• Key dates (anniversary, renewal)
• Stakeholders mentioned
• Competitive context
• Personal details shared

Risky Memory

Be careful with:

• Sensitive information
• Personal opinions expressed
• Complaints about individuals
• Confidential information shared

Consider:

• Do they expect you to remember this?
• Is it appropriate to reference?
• Could it backfire?

Memory Architecture

Data Model

Contact Profile:
  identity:
    - name, email, phone
    - company, role, timezone

  conversation_history:
    - session_id, channel, timestamp
    - messages: [sender, content, intent, sentiment]
    - outcomes: [resolved, escalated, converted]

  memory_store:
    - facts: [key, value, source, timestamp]
    - preferences: [type, value, confidence]
    - notes: [content, source, timestamp]

  state:
    - journey_stage
    - last_interaction
    - next_action
    - active_threads

Memory Operations

Store:

function storeMemory(contact_id, key, value, context) {
  memory = {
    key: key,
    value: value,
    source: context.channel,
    timestamp: now(),
    confidence: context.confidence
  }
  save(contact_id, memory)
}

Retrieve:

function getRelevantMemory(contact_id, current_context) {
  all_memory = load(contact_id)
  relevant = filter(all_memory, by_context)
  sorted = sort(relevant, by_relevance_and_recency)
  return top(sorted, n=5)
}

Update:

function updateMemory(contact_id, key, new_value) {
  existing = get(contact_id, key)
  if (new_value.timestamp > existing.timestamp) {
    update(contact_id, key, new_value)
  }
}

Using Memory in Conversations

Contextual Recall

Reference naturally: "You mentioned last time that budget approval was pending. Any update on that?"

Don't over-recall: Bad: "Hi Sarah! I remember you work at Acme Corp as a Sales Director, you're interested in our Pro plan, you have concerns about integration, your boss is named Tom, and you like to be contacted in the morning!"

Good: "Hi Sarah! Any update on the integration questions from last time?"

Memory-Informed Responses

Adjust based on history:

• Previous questions → Don't repeat
• Expressed preferences → Respect
• Past issues → Be sensitive

Example:

If (contact.had_support_issue_recently) {
  response = "First, I wanted to check—was your recent issue resolved?"
}

Memory Gaps

When information is uncertain: "I want to make sure I have this right—you mentioned being interested in [X]. Is that still the case?"

When information is missing: Don't pretend to know. Ask naturally.

Cross-Channel Memory

Unified View

Challenge: Customer interacts via chat, then email, then phone Solution: Single memory store, channel-aware

Implementation:

• Common contact ID across channels
• Merge interactions into unified timeline
• Channel-appropriate recall

Channel Context

Remember channel-specific details:

• Chat: Session-based, real-time
• Email: Thread-based, async
• SMS: Brief, immediate
• Phone: Rich, personal

Use appropriately: "I see we chatted on the website last week. Following up on that..."

Memory Maintenance

Data Freshness

Problem: Old data becomes wrong data

Solution:

• Timestamp all memories
• Decay confidence over time
• Re-verify important facts
• Archive old data

Example:

function shouldVerify(memory) {
  age = days_since(memory.timestamp)
  if (age > 90 && memory.type == "company_size") return true
  if (age > 30 && memory.type == "role") return true
  if (age > 7 && memory.type == "intent") return true
  return false
}

Conflict Resolution

When memories conflict:

• Prefer recent over old
• Prefer explicit over inferred
• Prefer high confidence over low
• Ask to verify when critical

Forgetting

When to forget:

• Explicit request
• Irrelevant information
• Privacy requirements
• Data retention policies

Privacy and Compliance

What to Remember

Ask yourself:

• Did they provide this information?
• Would they expect me to remember?
• Is it appropriate to use?
• Does it comply with regulations?

Data Rights

Support:

• Right to access (show what you know)
• Right to correct (fix wrong data)
• Right to delete (forget on request)
• Right to portability (export data)

Compliance

GDPR/CCPA considerations:

• Lawful basis for storing
• Purpose limitation
• Data minimization
• Storage limitation

Implementation Patterns

Memory Injection

Before generating response:

function generateResponse(message, contact_id) {
  // Get relevant memory
  memory = getRelevantMemory(contact_id, message.context)

  // Inject into prompt/context
  context = buildContext(message, memory)

  // Generate response
  response = llm.generate(context)

  // Store new memories
  extractAndStoreMemory(contact_id, message, response)

  return response
}

Memory Extraction

After each interaction:

function extractAndStoreMemory(contact_id, message, response) {
  // Extract facts from conversation
  facts = extractFacts(message.content)
  for (fact in facts) {
    storeMemory(contact_id, fact.key, fact.value, message.context)
  }

  // Update state
  updateState(contact_id, {
    last_interaction: now(),
    last_channel: message.channel,
    conversation_outcome: determineOutcome(response)
  })
}

Measuring Memory Effectiveness

Quality Metrics

Accuracy:

• Memory correctness rate
• Wrong recall rate
• Verification frequency

Coverage:

• % of interactions with relevant memory
• Memory utilization rate
• Gap identification

Experience Metrics

Engagement:

• Repeat question rate (lower = better)
• Customer satisfaction
• Conversation efficiency

Business:

• Conversion impact
• Support ticket reduction
• Relationship longevity

Common Mistakes

1. No Memory

Problem: Every conversation starts fresh Fix: Implement basic session and short-term memory

2. Too Much Memory

Problem: Overwhelming recall, creepy feeling Fix: Selective, context-appropriate recall

3. Stale Memory

Problem: Referencing outdated information Fix: Timestamp and refresh data

4. Wrong Memory

Problem: Mixing up customers, wrong facts Fix: Verification, conflict resolution

5. Siloed Memory

Problem: Each channel has separate memory Fix: Unified memory store

Questions to Ask

If you need more context:

1. What channels does your bot operate across?
2. How long are your typical customer relationships?
3. What CRM/data systems do you use?
4. What do customers complain about repeating?
5. What privacy/compliance requirements apply?

Related Skills

• personalization-at-scale: Using memory for personalization
• multi-channel-coordination: Memory across channels
• data-enrichment-integration: External data for memory
• conversational-flow-management: Context in conversations