CLT Memory

Cross-Layer Transcoder (CLT) memory extracts sparse features from model activations, providing interpretable representations of agent state.

How It Works

The CLTService extracts sparse features from inference activations using circuit-tracer:

state = await clt.extract_features(
    agent_id="critic",
    content="The risk model has issues...",
)
# state.features.active_indices — which features activated

Swarm Features

CLT features can be computed across a swarm to find consensus:

swarm_result = await clt.compute_swarm_features(domain="pension")
# swarm_result.consensus_features — features active across multiple agents

Integration with Topology

The TopologyService consumes CLT features to detect semantic attractors — regions in embedding space where agent attention converges:

CLTService → extract features → TopologyService → detect attractors → grid overlay

Qdrant Collections

CLT features are stored in dedicated Qdrant collections:

CLI Commands

# Extract CLT features
uv run gaius-cli --cmd "/clt extract" --format json

# CLT memory statistics
uv run gaius-cli --cmd "/clt stats" --format json