Dynamic Context

Learn how Dynamic Context enables more relevant, adaptive, and context-aware AI character interactions.

Introduction

Convai characters have two layers of context:

  • Character Level — backstory, personality, and behavioral rules. Configured in the Convai dashboard, persists across all sessions.

  • Session Level — information relevant only to the current interaction: who the player is, what just happened, the current game state.

Dynamic Context is a runtime feature of the Convai Unity SDK that lets your game push ephemeral, session-specific information to a Convai AI character during an active conversation. Unlike a character's fixed backstory configured in the Convai dashboard, Dynamic Context is injected at runtime and is temporary — it exists only for the duration of the current session.

Use it to bridge the gap between your game world and the AI: tell the character what just happened, who the player is, what they're carrying, where they are, or anything else that should influence the character's responses without being hardcoded into a prompt.


How It Works

When you call UpdateDynamicContext, the SDK serializes a context-update message and delivers it to the Convai backend over the active session's data channel. The backend receives the update and folds the new information into the character's active context window before processing the next user utterance.

The flow is:

Your Game Code


ConvaiRoomManager.UpdateDynamicContext(text, mode, runLlm)


Serialized context-update message dispatched over the session data channel


Convai Backend → LLM Context Window

This is a fire-and-forget, outbound-only operation. There is no direct response to the context update itself — the effect is visible in the character's next spoken reply.


Key Concepts

Before writing any code, understand these three parameters that control every UpdateDynamicContext call.

text

The content you want the character to be aware of. This is a plain string — it can be a sentence, a structured paragraph, or key-value data formatted however you prefer.

text is required for append and replace modes. When using reset mode, it is ignored.


mode

Controls how the new text is applied to the character's existing ephemeral context.

Mode
Behavior
When to Use

append

Adds text to any existing ephemeral context. Prior entries remain.

Progressive information — adding facts as the session unfolds.

replace

Discards all current ephemeral context and replaces it entirely with the new text.

A clean state transition — entering a new scene, starting a new chapter.

reset

Clears all runtime ephemeral context. text is ignored.

Resetting all accumulated runtime context.

Default: "append"


runLlm

Whether to trigger an LLM inference pass after the context update.

Value
Behavior

auto

The server decides based on its internal pipeline state. Recommended for most use cases.

true

Always triggers an LLM response immediately after the context update.

false

Updates the context silently — no LLM response is triggered.

Default: "auto"

Use "false" when you are batching multiple context updates and only want the LLM to respond after the final one. Use "true" when the context update itself should prompt a reaction from the character.


Accessing the API

Dynamic Context is exposed through ConvaiRoomManager, the central session manager for the Convai Unity SDK. It is accessible as a persistent singleton via ConvaiRoomManager.Instance.

Two distinct patterns exist depending on who owns the connection lifecycle.

Pattern A — Self-Contained

Use when the component is responsible for its own initialization.

The required readiness level depends on runLlm:

  • runLlm: "false" or "auto" — waiting for IsConnectedToRoom is sufficient. The context is injected silently and no immediate audio response is expected.

  • runLlm: "true" — the full audio pipeline must be ready before sending. See the note below.

circle-exclamation

Pattern B — Inject

Use when a session coordinator or parent system is responsible for providing the room manager. The component assumes the manager is valid when its public methods are called.

circle-info

UpdateDynamicContext returns false if the session is not connected. Always check the return value.


API Reference

Available on ConvaiRoomManager. Also callable via ConvaiNPC.UpdateDynamicContext(...), which delegates to ConvaiRoomManager.Instance internally.

Parameter
Type
Required
Default
Description

text

string

Yes (unless reset)

The context text to inject.

mode

string

No

"append"

How to apply the text: "append", "replace", or "reset".

runLlm

string

No

"auto"

Whether to trigger an LLM response: "true", "false", "auto".

Returns: true if the message was dispatched to the transport layer; false if the session is not connected (RTVIHandler not initialized) or if text is empty and mode is not "reset".


mode

Controls how the new text is applied to the character's existing ephemeral context.

Mode
Behavior
When to Use

append

Adds text to any existing ephemeral context. Prior entries remain.

Progressive information — adding facts as the session unfolds.

replace

Discards all current ephemeral context and replaces it entirely with the new text.

A clean state transition — entering a new scene, starting a new chapter.

reset

Clears all runtime ephemeral context. text is ignored.

Resetting all accumulated runtime context.

Default: "append"


Common Pitfalls

  • Calling before the session is connected UpdateDynamicContext returns false and silently drops the update. Use a coroutine or subscribe to ConvaiRoomManager.OnRoomConnectionSuccessful (a UnityEvent) to delay context injection until the session is live.

  • Using append indefinitely Every append expands the context window. In long sessions this degrades response coherence and increases latency. Use replace periodically, or adopt the DynamicContextManager pattern.

  • Sending text with reset mode text is ignored when mode is "reset". Pass null explicitly to make the intent clear.

  • Expecting a synchronous effect Dynamic Context is asynchronous end-to-end. The character will not know about the update until the next LLM inference pass. Do not design logic that assumes an instant effect.

  • Overusing runLlm: "true" Forcing an immediate response on every update creates conversational noise. Reserve it for moments where a character reaction is genuinely expected.


Quick Reference


Conclusion

Dynamic Context is the bridge between your game or application's live state and your AI character's understanding of it. The character's permanent configuration — personality, backstory, behavioral rules — stays stable in the Convai dashboard. Everything that belongs to this session, this user, and this moment flows through UpdateDynamicContext.

The API is intentionally simple: a single method with three parameters. The sophistication lives in how you use it — whether that is personalizing a conversation at session start, reacting to a threshold crossing mid-session, silently accumulating events until they are worth commenting on, or tracking competency gaps across an entire training session.

A few principles to carry forward:

  • Silent updates are the default posture. Use runLlm: "false" for state synchronization. Reserve runLlm: "true" for moments that genuinely warrant a character response.

  • Treat the context window as a resource. Unbounded append degrades quality over time. Use replace at natural boundaries — scene transitions, module changes, session resets.

  • Your application owns the state. The character receives it. Dynamic Context is the mechanism that enforces that separation cleanly.

circle-info

Need help? For questions, please visit the Convai Developer Forumarrow-up-right.

Last updated

Was this helpful?