Struct ReasoningConfig 

pub struct ReasoningConfig {

    pub enabled: bool,
    pub extract_provider: ProviderName,
    pub distill_provider: ProviderName,
    pub top_k: usize,
    pub store_limit: usize,
    pub max_messages: usize,
    pub max_message_chars: usize,
    pub context_budget_tokens: usize,
    pub min_messages: usize,
    pub extraction_timeout_secs: u64,
    pub distill_timeout_secs: u64,
    pub self_judge_window: usize,
    pub min_assistant_chars: usize,
}

ReasoningBank: distilled reasoning strategy memory configuration (#3342).

When enabled = true, each completed agent turn is evaluated by a self-judge LLM call. Successful and failed reasoning chains are compressed into short, generalizable strategy summaries. At context-build time, top-k strategies are retrieved by embedding similarity and injected into the prompt preamble.

All LLM work (self-judge, distillation) runs asynchronously — never on the turn thread.

Example

[memory.reasoning]
enabled = true
extract_provider = "fast"
distill_provider = "fast"
top_k = 3
store_limit = 1000

Fields

enabled: bool

Enable the reasoning-bank pipeline. Default: false.

extract_provider: ProviderName

Provider name from [[llm.providers]] for the self-judge step. Falls back to the primary provider when empty. Default: "".

distill_provider: ProviderName

Provider name from [[llm.providers]] for the distillation step. Falls back to the primary provider when empty. Default: "".

top_k: usize

Number of strategies retrieved per turn for context injection. Default: 3.

store_limit: usize

Maximum stored strategies; oldest unused are evicted when limit is reached. Default: 1000.

max_messages: usize

Maximum number of recent messages passed to the self-judge LLM. Default: 6.

max_message_chars: usize

Per-message content truncation limit (chars) before building the judge transcript. Default: 2000.

context_budget_tokens: usize

Maximum token budget for injected reasoning strategies in context. Default: 500.

min_messages: usize

Minimum number of messages required before self-judge fires. Default: 2.

extraction_timeout_secs: u64

Timeout in seconds for the self-judge LLM call. Default: 30.

distill_timeout_secs: u64

Timeout in seconds for the distillation LLM call. Default: 30.

self_judge_window: usize

Maximum number of recent messages passed to the self-judge evaluator. Narrowing to the last user+assistant pair improves classification accuracy. Default: 2.

min_assistant_chars: usize

Minimum characters in the assistant response to trigger self-judge. Short or trivial responses are skipped. Default: 50.

Trait Implementations

impl Clone for ReasoningConfig

fn clone(&self) -> ReasoningConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ReasoningConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for ReasoningConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for ReasoningConfig

where ReasoningConfig: Default,

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for ReasoningConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.