Skip to main content

Multi-Host Memory Topology

Run llmenv's memory backend (ICM) differently on different hosts: broadcast on one machine so trusted peers can share it, keep it loopback-only on another that shouldn't expose even an unauthenticated LAN service.

Config​

# ~/.config/llmenv/config.yaml

scope:
  host:
    - id: laptop
      match: { hostname: "laptop.local" }
      tags: [host-laptop, personal, work]
    - id: build-server
      match: { hostname: "build-server.internal" }
      tags: [host-build-server, server]

host:
  laptop:
    addr: "laptop.local"
  build-server:
    addr: "build-server.internal"

features:
  memory:
    # Laptop: on a trusted home/office LAN, broadcasts so other trusted
    # devices on that network can share the same memory context.
    - server_host: laptop
      listen_host: 0.0.0.0
      port: 9092
      when: [host-laptop, personal, work]

    # Build server: not on a fully trusted LAN. Loopback-only means the
    # backend is usable by agent sessions running on that host, but nothing
    # else on its network can reach the plaintext, unauthenticated port.
    - server_host: build-server
      listen_host: 127.0.0.1
      port: 9092
      when: [host-build-server]

How it works​

  • Each host scope emits its own tag (host-laptop, host-build-server), so only one features.memory entry is ever active on a given machine — the resolver would error if both matched the same scope simultaneously, but they can't, since the host scopes are mutually exclusive by hostname.
  • listen_host controls which interfaces the proxy binds, independent of server_host/addr (which is what other hosts use to reach it). 0.0.0.0 means any device that can route to that host can connect; 127.0.0.1 means only processes on that same machine can.
  • Both entries can point at the same port — they never run at the same time, since they're gated by disjoint host scopes.

See MCP & Memory → Security considerations before using 0.0.0.0 on anything but a network you actually trust — the memory backend has no auth and no TLS; reachability is the entire access model.

Verify​

# On each host:
llmenv context
# shows which features.memory entry is active and its resolved listen_host