Crate talos_evolution 

Talos Evolution — self-evolution engine for agent behavior adaptation.

Implements a 4-phase learning loop per ADR-001:

1. Observe: Capture signals during agent execution
2. Extract: Identify patterns from observations
3. Store: Persist patterns with confidence scores
4. Apply: Inject high-confidence patterns into system prompt

Re-exports

pub use hook::EvolutionHookHandler;

Modules

adapter

BehaviorAdapter — injects bounded, advisory patterns into system prompt.

extractor

PatternExtractor — extracts patterns from observations.

hook

Evolution HookHandler — wires the I008 self-evolution engine (ADR-001) into any agent run path via the I009 hook system (see talos_plugin::handler::HookHandler).

observer

TurnObserver — captures signals during agent execution.

store

KnowledgeStore — SQLite persistence for observations and patterns.

Structs

Conflict

A conflict between two patterns.

EvolutionConfig

Configuration for the evolution engine.

Observation

An observation captured from a single turn.

Pattern

A pattern extracted from multiple observations.

Signal

A single learning signal captured during agent execution.

ToolUsage

Record of a tool invocation during a turn.

TurnObservation

Per-turn observation that aggregates multiple signals with turn-level metadata.

Enums

EvolutionError

Error type for the evolution crate.

SignalKind

The kind of learning signal captured during agent execution. Four base variants per the MenteDB cognitive-feedback blueprint (docs/reference/REFERENCE-PROJECTS.md §17).

SignalType

A signal captured during agent execution.

StoreError

Errors originating from the SQLite knowledge store.

TurnOutcome

Outcome of a single agent turn.

Functions

compute_content_hash

Compute a normalized fingerprint for content-based dedup. Format: “{category}|{first 1KB of instruction}” hashed via DefaultHasher.

Type Aliases

EvolutionResult

Result type for evolution operations.