Memory

Cortex provides three types of memory for maintaining context across interactions:

Memory types

Conversation memory

Stores the full message history between a user and an agent. Scoped by agent ID and tenant ID.

Working memory

Temporary key-value storage for a single run. Used for intermediate reasoning state. Cleared when the run completes.

Summary memory

Compressed summaries of past conversations. Used to provide context when the full conversation history is too long.

Message

type Message struct {
    Role      string         // "user", "assistant", "system", "tool"
    Content   string
    ToolCalls []any
    Metadata  map[string]any
    Timestamp time.Time
}

Store interface

type Store interface {
    // Conversation memory
    SaveConversation(ctx context.Context, agentID id.AgentID, tenantID string, messages []Message) error
    LoadConversation(ctx context.Context, agentID id.AgentID, tenantID string, limit int) ([]Message, error)
    ClearConversation(ctx context.Context, agentID id.AgentID, tenantID string) error

    // Working memory
    SaveWorking(ctx context.Context, runID id.AgentRunID, key string, value any) error
    LoadWorking(ctx context.Context, runID id.AgentRunID, key string) (any, error)
    ClearWorking(ctx context.Context, runID id.AgentRunID) error

    // Summary memory
    SaveSummary(ctx context.Context, agentID id.AgentID, tenantID string, summary string) error
    LoadSummaries(ctx context.Context, agentID id.AgentID, tenantID string) ([]string, error)
}

The memory store has 8 methods across three memory types. All conversation and summary operations are scoped by agent ID and tenant ID.

API routes

Method	Path	Description
`GET`	`/cortex/agents/{name}/memory`	Load conversation history
`DELETE`	`/cortex/agents/{name}/memory`	Clear conversation history