Core Concepts

Memory Layers

The 10-layer cognitive architecture that powers ekkOS.

ekkOS organizes memory into 10 specialized layers, each serving a distinct cognitive function. This mirrors how human memory works — from short-term working memory to long-term semantic knowledge. Information flows upward through the layers, becoming more abstract and permanent over time.

The 10 Layers

Working Memory

24 hours

Immediate context

Raw chat messages from current and recent sessions. This is the scratchpad where new information first lands before being processed into higher layers.

Current conversation messagesRecent file edits discussedCommands executed

Episodic Memory

30 days

Significant events

Complete problem-solution cycles extracted from working memory. Each episode captures a meaningful interaction with context, actions, and outcomes.

Bug fix sessionsFeature implementationsConfiguration changes

Semantic Memory

Permanent

Factual knowledge

Compressed, factual knowledge distilled from episodes. Stored as vector embeddings for semantic search. Facts without the narrative.

"Project uses PostgreSQL""Auth is handled by Supabase""API routes are in /api/v1"

Pattern Memory

Permanent

Reusable solutions

Problem-solution pairs that work. Each pattern has a confidence score that evolves based on application outcomes. The core of learning.

Authentication patternsError handling approachesDatabase query optimizations

Procedural Memory

Permanent

Multi-step workflows

Step-by-step procedures for complex tasks. Useful when order matters and multiple actions must be coordinated.

Deployment workflowsDatabase migration stepsCI/CD pipelines

Collective Memory

7 days

Cross-agent learning

Knowledge shared between all AI agents connected to your account. What Claude Code learns, Cursor can use.

Patterns discovered in Cursor available to Claude CodeCross-tool context continuity

Meta Memory

Permanent

Self-awareness

The system monitoring itself. Tracks pattern effectiveness, identifies drift, and triggers consolidation when needed.

Pattern success ratesMemory layer healthConfidence score trends

Codebase Memory

Permanent

Code understanding

Vector embeddings of your codebase. Enables semantic search like "find the authentication logic" to locate relevant code.

Function embeddingsFile summariesArchitectural patterns

Directives

Permanent

Behavioral rules

Explicit MUST/NEVER/PREFER/AVOID rules with priority levels. High-priority directives override other knowledge.

"MUST use TypeScript""NEVER commit secrets""PREFER functional components"

Conflict Resolution

Permanent

Decision audit

Logs of how conflicting information was resolved. Maintains audit trail for understanding AI decisions.

Pattern A vs Pattern B decisionsDirective conflictsConfidence tiebreakers

Data Flow

Chat Message
    ↓
┌─────────────────┐
│ L1: Working     │ ← Raw messages (24h TTL)
└────────┬────────┘
         ↓ (every 5 min)
┌─────────────────┐
│ L2: Episodic    │ ← Extracted episodes (30d TTL)
└────────┬────────┘
         ↓ (compression)
┌─────────────────┐
│ L3: Semantic    │ ← Factual knowledge (permanent)
└────────┬────────┘
         ↓ (pattern discovery)
┌─────────────────┐
│ L4: Pattern     │ ← Reusable solutions (permanent)
└─────────────────┘