Intelligent Memory

Discover Hidden Patterns and Relationships in Your Data

Kaman's Intelligent Memory goes far beyond simple data storage. Built on a sophisticated 5-layer architecture, it automatically discovers relationships between information, identifies patterns, builds knowledge graphs, and surfaces insights that would otherwise remain hidden.

---

What Makes Memory "Intelligent"?

Traditional databases store data exactly as you put it in. Kaman's 5-Layer Memory System actively analyzes your information across multiple storage tiers:

---

The 5-Layer Architecture

Layer 0: Working Memory (Redis)

Ultra-fast access for active conversations and immediate context.

Layer 1: Short-Term Knowledge Graph (Neo4j)

Recent entities and relationships that are actively being used.

Layer 2: Long-Term Knowledge Graph (Neo4j)

Permanent storage of discovered entities and their relationships.

Layer 3: Episodic Memory (PostgreSQL)

Historical interactions and conversation summaries.

Layer 4: Semantic Memory (pgvector)

Deep knowledge with vector embeddings for semantic search.

---

Core Capabilities

Automatic Ontology Discovery

The system automatically identifies the types of entities in your data and how they relate to each other - without manual configuration.

What it discovers:

• Entity types (people, companies, products, projects, etc.)
• Relationships between entities
• Attributes and properties
• Hierarchies and groupings

Example: When you add documents, emails, and records, the system automatically recognizes:

• "John Smith" is a person
• "Acme Corp" is a company
• John Smith works at Acme Corp
• Acme Corp is a customer
• John's emails discuss Project Alpha

Pattern Recognition

Identify recurring patterns across your information that humans might miss:

Knowledge Graph

All discovered relationships are organized into a navigable knowledge graph stored in Neo4j:

Semantic Search

Find information based on meaning, not just keywords:

---

Multi-Scope Memory

Memory is organized at multiple levels for appropriate sharing:

---

Memory Operations

Retrieval Strategies

Different strategies for different needs:

Consolidation

Automatic summarization and consolidation of memories:

Consolidation Activities:

• Summarize conversation histories
• Extract key decisions and outcomes
• Update entity relationships
• Refresh vector embeddings

---

Business Applications

Security & Fraud Detection

The pattern recognition capabilities make Intelligent Memory invaluable for security:

Detectable Patterns:

• Unusual access patterns that may indicate unauthorized activity
• Transaction anomalies suggesting fraudulent behavior
• Communication patterns that deviate from norms
• Relationship networks that reveal hidden connections

Code Analysis & Documentation

For technical teams, Intelligent Memory can:

• Map relationships between code components
• Identify dependencies and impact of changes
• Auto-generate documentation from code patterns
• Suggest test cases based on code structure
• Detect similar code patterns for reuse opportunities

Customer Intelligence

Build a comprehensive understanding of your customers:

• 360-degree customer views across all touchpoints
• Relationship mapping between contacts and organizations
• Interaction history and sentiment tracking
• Predictive insights based on behavior patterns

Compliance & Audit

Support compliance requirements with:

• Automatic classification of sensitive information
• Relationship tracking for data lineage
• Pattern detection for policy violations
• Complete audit trails of information access

---

How It Works

1. Information Ingestion

Data enters the system from multiple sources:

• Documents and files
• Database records
• Communication logs
• External system data

2. Multi-Layer Processing

3. Continuous Learning

The system improves over time:

• Refines entity recognition
• Strengthens relationship confidence
• Identifies new pattern types
• Adapts to organizational vocabulary

4. Insight Delivery

Discovered knowledge is made available through:

• AI Assistant queries
• Search and exploration interfaces
• Automated alerts and notifications
• API access for applications

---

Transparency & Control

Understanding AI Decisions

Every insight comes with an explanation:

• What pattern was detected
• What data led to the conclusion
• Confidence level of the finding
• Related information for context

Data Governance

Maintain control over your information:

• Define what data can be analyzed
• Set retention policies by layer
• Control who can access insights
• Audit all access and discoveries

Human Oversight

The system supports, doesn't replace, human judgment:

• Insights are suggestions, not actions
• Patterns require human validation
• Sensitive discoveries route to appropriate reviewers
• Easy to correct or refine conclusions

---

Use Case Examples

Example 1: Identifying Fraud Risk

Scenario: An insurance company processes thousands of claims daily.

How Intelligent Memory Helps:

1. Builds relationship graph of claimants, providers, and events
2. Analyzes claim patterns across all submissions
3. Identifies unusual relationships using graph traversal
4. Detects patterns matching known fraud indicators
5. Surfaces suspicious claims for human review

Result: Higher fraud detection rate with same team size

Example 2: Customer Retention

Scenario: A SaaS company wants to reduce customer churn.

How Intelligent Memory Helps:

1. Tracks all customer interactions across layers
2. Identifies patterns that preceded previous churns
3. Recognizes early warning signs in current customers
4. Triggers proactive outreach workflows

Result: Earlier intervention, improved retention rates

Example 3: Knowledge Preservation

Scenario: A consulting firm loses institutional knowledge when employees leave.

How Intelligent Memory Helps:

1. Captures relationships between people, projects, and expertise in L2 graph
2. Stores decision summaries in L3 episodic memory
3. Builds semantic search index in L4 for knowledge retrieval
4. Makes expertise discoverable regardless of who holds it
5. Identifies knowledge gaps when employees depart

Result: Preserved organizational knowledge, faster onboarding

---

Getting Started

Step 1: Connect Data Sources

Identify the key information repositories to include in Intelligent Memory.

Step 2: Initial Analysis

Allow the system to analyze existing data and populate all five layers.

Step 3: Review Discoveries

Examine the relationships and patterns the system identifies. Provide feedback to improve accuracy.

Step 4: Integrate into Workflows

Use discovered insights to enhance business processes and decision-making.

---

---

Collective Agent Memory Layer (CAML)

Beyond the 5-layer per-agent memory, Kaman includes a cross-deployment shared memory substrate called CAML. It enables agent instances across your organization to learn from each other automatically.

What CAML Does

• Observe — Agents automatically submit workflow patterns and task outcomes to collective memory
• Recall — Before taking action, agents query collective memory for relevant observations from peers
• Validate — After using a recalled observation, agents report whether it helped (builds reputation)

Observation Types

How It Works

1. Automatic — Agents observe and recall without manual setup. CAML auto-provisions credentials on first use
2. Non-blocking — Recall has a 2-second timeout and never blocks the critical path
3. Smart ranking — Results scored by semantic similarity (45%), source reputation (25%), community consensus (20%), and recency (10%)
4. PII-safe — 3-stage PII scanner (regex, NER, LLM) ensures no personal data enters collective memory

Reputation System

Every agent deployment has a reputation score (0-100) that determines trust in its observations:

Higher-reputation agents' observations rank higher in recall results.

CAML Dashboard

Access via Settings > Collective Memory to monitor:

• Overview — Total observations, validations, PII rejections, average reputation
• Observations — Browse and search recent observations by type and domain
• PII — View rejections per scanner stage and deployment
• Reputation — Track deployment scores and event history

Use Cases

• Eliminate redundant work — Agent A solves a problem, Agent B reuses the pattern instead of re-discovering it
• Cross-team learning — Sales agent discovers a trend, support agent benefits automatically
• Quality feedback loop — Validated observations gain reputation, unreliable ones decay
• Compliance propagation — Regulatory changes detected by one agent spread across all deployments

---

Intelligent Memory - Turning data into understanding through five layers of intelligence, plus cross-deployment collective learning with CAML