Skip to content

Kaizen bridge arch

Kaizen Bridge Architecture

Overview

Kaizen Bridge implements Cache-to-Cache (C2C) semantic federation, enabling direct, low-latency semantic transfer between agents (e.g., AUREA ↔ HERMES) by projecting one model's KV/hidden states into another's latent space.

Why Cache-to-Cache?

Traditional multi-agent systems send text back and forth, requiring each model to: - Decode → Interpret → Generate → Re-encode

This burns latency and loses nuance (the "semantic compression" problem).

C2C bypasses that by letting one model send its KV-Cache vectors — basically its live thought state — directly into another's network, with a projection layer to align meaning.

Benefits

  1. Latency — 2× speedup across federated agent loops
  2. Integrity Preservation — no semantic loss during transfer
  3. Scalability — allows direct cognitive mesh among agents without text bottlenecks

Architecture

HERMES Agent → KV Extract → Broker → Projector → AUREA Agent
                ↓                                    ↓
            Ledger Attestation              State Buffer

Components

  1. Broker (services/kaizen-bridge/broker/)
  2. Express.js service handling relay requests
  3. Routes packets to appropriate projector
  4. Handles HMAC signatures and attestation

  5. Emits SSE events for dashboard visualization

  6. Projector (services/kaizen-bridge/projector/)

  7. Python FastAPI service
  8. MLP-based projection models

  9. Transforms KV cache tensors between agent pairs

  10. Agent Stubs (services/agents/{hermes,aurea}/)

  11. KV cache extraction (or simulation)
  12. Bridge client integration
  13. State buffer and blending

Trust Model

  • Signed Relay: HMAC signatures on broker requests
  • Ledger Attestations: Each relay writes {from,to,projector,input_hash,output_hash,ts} to civic-ledger
  • GI Floors: Enforce minimum GI scores (default: 0.95) for relay eligibility

Phased Rollout

  1. Shadow Mode: Relay + projection, target still generates text from projected state
  2. Hybrid: Target blends projected state with its own cache
  3. Direct: Target accepts projected vectors for downstream layers

Security

  • HMAC authentication on broker endpoints
  • Payload size & rate limits
  • No raw user data; only numerical tensors encoded in base64
  • Ledger attestation for audit trail

Future Enhancements

  1. Extractor plugins for each provider (OpenAI, Anthropic, local) to standardize KVCachePacket
  2. Learned alignment (train the MLP on parallel traces; add checkpoint save/load)
  3. GRPC path for lower overhead; optional quantization (int8) to cut bandwidth
  4. Privacy mode: on-device projection for sensitive agents (no network hop)
  5. GI-aware routing: deny relay if GI < floor or pair not whitelisted