Operator wraps the entire Claude agent lifecycle in a hook-driven architecture that remembers everything, routes intelligently, and improves autonomously. The longer you use it, the better it gets.
$ curl -fsSL https://operator.lemay.app/install.sh | bash
click to copy
29
CDI Memory Tools
12
Lifecycle Hooks
9
Hook Events
5
Learning Systems
10
Router Layers
Capabilities
Everything an agent should be
Operator is not a tool. It is an operating system layer that makes Claude remember, learn, protect, predict, and improve.
M
Persistent Memory
Every conversation, decision, error, and tool result is stored in the CDI and survives restarts, compaction, and session changes. Cross-session queries bring prior knowledge into every new task.
R
Intelligence Router
Ten-layer classification on every prompt: effort, model selection, token budget, subsystem activation, context strategy, and more. The router reads learned overrides and improves with every consolidation cycle.
H
Hook Pipeline
Twelve hooks across nine lifecycle events. Every interaction is observed, validated, captured, and enforced. The triple-gate Stop hook ensures nothing completes without memory storage and routing feedback.
S
Security Boundary
bash-guard validates shell commands against config-driven blocked patterns. evolution-guard enforces immutable infrastructure paths. Every blocked action is logged to a security audit trail.
L
Self-Learning
Five learning systems operate continuously: memory consolidation, router self-tuning, predictive dispatch, skill evolution, and subagent memory sharing. The agent compounds intelligence across every session.
C
Centralized Config
Every threshold, pattern, limit, and feature flag in a single config.json. Shared library with 10 functions sourced by all scripts. Error resilience with safe fallbacks on every hook.
Hook Pipeline
Every interaction, observed
From session start to session end, twelve hooks monitor, validate, capture, and enforce across nine lifecycle events.
SessionStart
→
UserPromptSubmit
→
PreToolUse
→
PostToolUse
→
Stop
→
SessionEnd
PreCompact
SubagentStop
Notification
Status bar feedback on every firing — you always know Operator is working
Five systems operate silently across sessions. The agent doesn't just remember — it recognizes patterns, tunes its own routing, anticipates needs, and rewrites underperforming skills.
01
Memory Consolidation
Seven pattern detectors run inference over accumulated CDI: error recurrence, decision convergence, tool chain optimization, user preference extraction, environment stability, recovery synthesis, and temporal patterns. Raw memory becomes learned knowledge.
02
Self-Tuning Router
Every routing classification is tracked against actual outcomes. Misclassifications are mined for patterns. Learned overrides are written to router-learned.json and applied before static classifiers on the next prompt.
03
Predictive Dispatch
Temporal and sequential patterns feed anticipatory actions. The agent prepares what you need before you ask. Risk-tiered: low-risk auto-executes, medium-risk suggests, high-risk surfaces awareness only.
04
Skill Evolution
Performance is tracked on every invocation. Skills with failure rate above 30% are candidates for autonomous rewrite. Every modification is evidence-based, backed up, auditable, and reversible.
05
Subagent Memory Sharing
When spawning Agent subagents, CDI context is queried and injected into the subagent prompt. The SubagentStop hook captures results with parent-child linking. Multi-agent coordination through shared persistent memory — subagents benefit from every prior session.
Architecture
The full picture
Context window as L1 cache. CDI as permanent memory. Hooks as the nervous system. Learning systems as the cortex.