TokenCounter

TokenCounter = (text) => number

Pluggable token counting interface.

Used by memory.compaction() and model.router() to estimate token counts. The default implementation uses a chars / 4 heuristic (~85% accuracy). Users can plug in tokenx (~95%) or js-tiktoken (100%) for better accuracy.

Parameters

text

string

Returns

number

Example

// Default (built-in, zero deps):
memory.compaction({ maxTokens: 8192 })

// Better estimation (user installs tokenx):
import { estimateTokenCount } from "tokenx"
memory.compaction({ maxTokens: 8192, tokenCounter: (text) => estimateTokenCount(text) })

// Exact counting (user installs js-tiktoken):
import { encodingForModel } from "js-tiktoken"
const enc = encodingForModel("gpt-4o")
memory.compaction({ maxTokens: 8192, tokenCounter: (text) => enc.encode(text).length })