Skip to main content

​
User Preferences & Personalization

Every user can customize their AgentFlow experience — from which LLM model the agent uses, to the tone and style of responses, to which capabilities are enabled.

​
Settings API

​
Get current settings

GET /api/v1/settings
Returns the merged settings (user overrides + platform defaults).

​
Update settings

PATCH /api/v1/settings
{
  "personalization": {
    "baseStyleTone": "concise and professional",
    "customInstructions": "Always include action items at the end",
    "aboutUser": "I'm a VP of Customer Success managing a team of 12 CSMs"
  },
  "model": {
    "selectedModel": "anthropic/claude-sonnet-4"
  }
}

​
Reset to defaults

POST /api/v1/settings/reset

​
Get platform defaults

GET /api/v1/settings/defaults

​
Settings categories

​
Personalization

SettingTypeDescription
baseStyleTonestringResponse style (e.g., “concise”, “detailed”, “casual”)
customInstructionsstringPersistent instructions applied to every conversation
aboutUserstringContext about the user (role, team, preferences)
Personalization is injected into the agent’s context only when it’s relevant — the agent sees the user’s profile but is instructed to use it naturally, not robotically.

​
Model selection

SettingTypeDescription
selectedModelstringDefault model for all agent interactions
pinnedDefaultModelstringFallback model when no preference set
backgroundModelstringModel for background operations (title generation, memory)

​
Capabilities

SettingTypeDescription
enablePlanningbooleanMulti-step task decomposition
enableReflectionbooleanSelf-evaluation of responses

​
Generation parameters

SettingTypeDescription
reasoningEffortstringDefault reasoning depth (low, medium, high)
reasoningSummarystringReasoning output mode: "auto", "concise", "detailed"

​
Behavior

SettingTypeDescription
streamingbooleanEnable response streaming (default: true)
followUpEnabledbooleanSuggest follow-up questions
autocomplete.enabledbooleanEnable input autocomplete suggestions

​
Adaptive Memory

Beyond explicit settings, AgentFlow can learn user preferences through sleep-time memory. After conversation activity settles, a background process analyzes the completed conversation and distills durable facts and preferences into persistent memory blocks. This means the agent gradually learns things like:
  • “This user prefers bullet-point summaries”
  • “This user manages the Acme and Globex accounts”
  • “This user always wants meeting prep to include recent email context”
Memory is stored per user and tenant, with optional agent-specific blocks. Core memory is injected into future prompts; archival memory is recalled by query when relevant. See Memory for the full model and SDK: Memory for Python examples.