Memory System

Kage uses a two-tier memory system to store and share context between agents.

Two-Tier Architecture

Working Memory

• In-memory, fast access
• Current session context
• Cleared on agent termination

Long-Term Memory

• Persisted to disk (append-only logs)
• Discoveries and learned patterns
• Survives restarts

Memory Types

Kage stores different types of memories:

pub enum MemoryEntry {
    // Code understanding
    FileDiscovered { path, summary, structure },
    PatternLearned { pattern, examples, confidence },
    DependencyMapped { from, to, relationship },

    // Task context
    ErrorEncountered { error, resolution, worked },
    DecisionMade { decision, reasoning, alternatives },
    TaskCompleted { task_id, summary, artifacts },

    // Collaboration
    InsightShared { topic, content, from_agent },
    QuestionAsked { question, answer, from_agent },
}

Memory Scopes

Memories can be scoped to different levels:

• agent - Private to one agent
• namespace - Shared within namespace
• global - Available to all agents (opt-in)

Context Injection

Inject context from one agent to another:

kage context inject <target-agent> --from <source-agent>

Filter by type or topic:

kage context inject agent-2 --from agent-1 --type patterns
kage context inject agent-2 --from agent-1 --topic "auth"