Struct MemCotConfig 

pub struct MemCotConfig {
    pub enabled: bool,
    pub distill_provider: ProviderName,
    pub distill_timeout_secs: u64,
    pub min_assistant_chars: usize,
    pub min_distill_interval_secs: u64,
    pub max_distills_per_session: u64,
    pub max_state_chars: usize,
    pub recall_view: RecallViewConfig,
    pub zoom_out_neighbor_cap: usize,
    pub fast_tier_models: Vec<String>,
}

MemCoT semantic-state distillation configuration.

When enabled = true, the agent maintains a short rolling “semantic state” buffer summarizing conceptual progress across turns. This buffer is injected into graph recall queries to improve retrieval relevance.

All LLM work (distillation) runs asynchronously — never on the turn thread. When enabled = false, this is a complete no-op: no allocation, no LLM calls.

Config example

[memory.memcot]
enabled = true
distill_provider = "fast"
distill_timeout_secs = 5
min_assistant_chars = 200
min_distill_interval_secs = 30
max_distills_per_session = 50
max_state_chars = 800
recall_view = "head"

Fields

enabled: bool

Enable the MemCoT semantic state pipeline. Default: false.

When false, the accumulator is never allocated and no LLM calls are made.

distill_provider: ProviderName

Provider name from [[llm.providers]] for distillation.

Must reference a fast-tier provider (e.g. gpt-4o-mini, qwen3:8b). A startup warning is emitted when the resolved model does not look fast-tier. Falls back to the primary provider when empty. Default: "".

distill_timeout_secs: u64

Timeout in seconds for each distillation LLM call. Default: 5.

min_assistant_chars: usize

Minimum characters in the assistant response to trigger distillation. Short or trivial replies are skipped. Default: 200.

min_distill_interval_secs: u64

Minimum elapsed seconds between successive distillation spawns. Default: 30.

Prevents runaway costs on long sessions with rapid turns. Clearing /new resets this counter.

max_distills_per_session: u64

Maximum distillation spawns per conversation session. Default: 50.

Once this cap is reached the accumulator stops distilling for the rest of the session. Counter is reset when the user sends /new.

max_state_chars: usize

Maximum characters for the semantic state buffer (UTF-8 char boundary truncation). Default: 800.

recall_view: RecallViewConfig

Recall view applied when MemCoT is active. Default: Head.

• head: standard retrieval, no enrichment (suitable for low-latency setups).
• zoom_in: adds source-message provenance to each returned fact.
• zoom_out: expands 1-hop neighbors per returned fact.

TODO(F3): add a per-call override parameter on recall_graph_view.

zoom_out_neighbor_cap: usize

Maximum 1-hop neighbor facts per head fact in zoom_out view. Default: 3.

fast_tier_models: Vec<String>

Optional model name allowlist for the fast-tier soft validator (lowercase substring match). Empty (default) → falls back to the built-in FAST_TIER_MODEL_HINTS list.

Trait Implementations

impl Clone for MemCotConfig

fn clone(&self) -> MemCotConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for MemCotConfig

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

Formats the value using the given formatter.

impl Default for MemCotConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for MemCotConfig

where MemCotConfig: 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 MemCotConfig

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

where __S: Serializer,

Serialize this value into the given Serde serializer.