Token Budget

Budget Formula

// Token budget calculation
HistoryBudget = MaxWorkingMemoryTokens
              - SystemPromptTokens
              - ProcedureTokens       (if matched)
              - RetrievedKnowledgeTokens
              - EpisodeSnippetTokens
              - CurrentMessageTokens
              - OutputReserve         // tokens reserved for agent response

// Example with Claude and a large knowledge base:
// MaxWorkingMemoryTokens:   150,000
// SystemPromptTokens:        -1,200
// ProcedureTokens:             -300
// RetrievedKnowledgeTokens:  -1,500
// EpisodeSnippetTokens:        -400
// CurrentMessageTokens:        -100
// OutputReserve:             -8,192
// ─────────────────────────────────
// HistoryBudget:            138,308  tokens available for message history

Token Counting Approaches

// Default approximate token counter
public class ApproximateTokenCounter : ITokenCounter
{
    // Rough approximation: ~4 characters per token for English text
    public int Count(string text) => Math.Max(1, text.Length / 4);

    public int CountMessages(IReadOnlyList<LLMMessage> messages) =>
        messages.Sum(m => Count(m.Content) + 4);  // +4 for role overhead
}

Budget Configuration

// MemoryConfig budget settings
public class MemoryConfig
{
    // Context window settings
    public int MaxWorkingMemoryTokens { get; set; } = 100_000;
    public int OutputReserve { get; set; } = 4_096;      // reserved for response
    public int MinHistoryTokens { get; set; } = 4_000;   // always keep at least this

    // Section limits
    public int MaxKnowledgeTokens { get; set; } = 3_000;  // cap semantic memory size
    public int MaxEpisodeTokens { get; set; } = 1_000;    // cap episodic snippet size
}