Module reasoning 

ReasoningBank: distilled reasoning strategy memory (#3342).

After each completed agent turn a three-stage async pipeline runs off the hot path:

1. Self-judge (run_self_judge) — a fast LLM evaluates success/failure and extracts the key reasoning steps.
2. Distillation (distill_strategy) — a strategy summary (≤ 3 sentences) is generated from the reasoning chain, capturing the transferable principle.
3. Storage (ReasoningMemory::insert) — the summary is written to SQLite and, when Qdrant is available, embedded and indexed for vector retrieval.

At context-build time ReasoningMemory::retrieve_by_embedding fetches top-k strategies by embedding similarity. The caller (in zeph-context) calls ReasoningMemory::mark_used only for strategies actually injected into the prompt, after budget truncation (C4 split from architect plan).

LRU eviction

ReasoningMemory::evict_lru protects rows with use_count > HOT_STRATEGY_USE_COUNT (default 10) from normal eviction. When all rows are hot and the table exceeds 2 × store_limit, a forced eviction pass deletes the oldest rows unconditionally and emits a warn! so operators can tune store_limit upward.

LRU eviction race note

Two concurrent turns may race on the count check in evict_lru. Either both evict (over-eviction by at most top_k rows) or neither. This is acceptable for MVP — the table remains bounded.

Structs

OutcomeParseError

Error returned when parsing an Outcome from a string fails.

ProcessTurnConfig

Configuration for the process_turn extraction pipeline.

ReasoningMemory

SQLite-backed store for distilled reasoning strategies.

ReasoningStrategy

A distilled reasoning strategy row from the reasoning_strategies table.

SelfJudgeOutcome

Parsed response from the self-judge LLM call.

Enums

Outcome

Outcome of a reasoning strategy: whether the agent succeeded or failed.

Constants

REASONING_COLLECTION

Qdrant collection name used for reasoning-strategy embeddings.

Functions

distill_strategy

Run the distillation step.

process_turn

Run the full extraction pipeline for a single turn.

run_self_judge

Run the self-judge step against a turn’s message tail.