feat(ai-memory): central agent memory server on m3-hermes

Server: services.ai-memory (upstream NixOS module, llm-agents build) on
m3-hermes, bound to the Netbird range, LLM consolidation via z.ai GLM,
one global cross-project profile, auto-improve proposals gated behind
approval. Traefik on m3-atlas publishes it as memory.m3ta.dev.

Clients: per-host aim_ keys (user "sascha", labels kratos / hermes /
az-lt-nix) as agenix secrets. m3-kratos enables m3ta-home's
coding.aiMemory with the AZ checkouts routed to the "az" workspace.

Hermes: lifecycle hooks, MCP bridge and skills wired declaratively, a
home-workspace marker ("home"/"assistant"), and Honcho replaced by the
built-in notes so ai-memory is the single shared memory.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
m3tam3reandClaude Fable 5.1 committed 2026-10-25 13:01:05 +01:00
1 parent 35a0a45977
commit f142cc0721
15 files changed
+274 -12

No files matched your search

+7
View File
@@ -10,6 +10,13 @@
hermes-api-server-key = {
file = ../../secrets/hermes-api-server-key.age;
};
ai-memory-server-env = {
file = ../../secrets/ai-memory-server-env.age;
};
ai-memory-hermes-token = {
file = ../../secrets/ai-memory-hermes-token.age;
owner = "hermes";
};
};
};
}
+52
View File
@@ -0,0 +1,52 @@
{
config,
pkgs,
...
}: let
# Netbird mesh VPN range — the server is only reachable from mesh peers.
# m3-atlas Traefik proxies memory.m3ta.dev to this port over Netbird.
netbirdRange = "100.64.0.0/16";
port = config.m3ta.ports.get "ai-memory";
in {
# ── ai-memory server ──────────────────────────────────────────────────
# Shared agent memory for all hosts (private + work), see
# https://aimemory.sh. Clients: m3ta-home `coding.aiMemory`.
#
# Flow: Agent → memory.m3ta.dev (TLS via m3-atlas Traefik) → Netbird → :${toString port}
# Hermes on this host talks to 127.0.0.1 directly.
services.ai-memory = {
enable = true;
# Same build as the clients (m3ta-home uses llm-agents.ai-memory),
# prebuilt via the numtide cache instead of compiling the flake.
package = pkgs.llm-agents.ai-memory;
bind = "0.0.0.0";
inherit port;
enableWeb = true;
# Root bearer, token pepper, recovery token, cookie/proxy settings and
# the LLM provider key. Never put any of these into `settings`.
ageSecret = "ai-memory-server-env";
settings = {
allowed_hosts = ["memory.m3ta.dev" "localhost" "127.0.0.1" "::1"];
# One operator, several workspaces (private, az, home): keep one
# cross-project profile for all of them. With database users the
# profile is off unless enabled explicitly.
profile = {
enabled = true;
share = "global";
};
# Shared server: auto-improvement proposals wait for approval.
auto_improve.require_approval = true;
};
};
networking.firewall = {
extraCommands = ''
# Allow ai-memory only from Netbird mesh VPN
ip46tables -A nixos-fw -p tcp --dport ${toString port} -s ${netbirdRange} -j nixos-fw-accept
'';
extraStopCommands = ''
ip46tables -D nixos-fw -p tcp --dport ${toString port} -s ${netbirdRange} -j nixos-fw-accept 2>/dev/null || true
'';
};
}
+1
View File
@@ -3,6 +3,7 @@
./hermes-agent.nix
./herdr-sandbox.nix
./hermes-dashboard.nix
./ai-memory.nix
./netbird.nix
];
}
+73 -2
View File
@@ -13,12 +13,45 @@
# Same interpreter env the upstream merge script uses, so no extra
# closure cost — the derivation is shared.
yamlPython = pkgs.python3.withPackages (ps: [ps.pyyaml]);
# ── ai-memory client (server: ./ai-memory.nix, same host) ─────────────
# Hooks and the MCP bridge read the bearer from [auth].bearer_token_file
# in the data dir's config.toml, so neither config.yaml nor the process
# environment ever carries the token.
aiMemory = "${config.services.ai-memory.package}/bin/ai-memory";
aiMemoryUrl = "http://127.0.0.1:${toString (config.m3ta.ports.get "ai-memory")}";
aiMemoryDataDir = "/var/lib/hermes/.local/share/ai-memory";
aiMemoryHook = event: [
{
command = "${aiMemory} --data-dir ${aiMemoryDataDir} hook --event ${event} --agent hermes --server-url ${aiMemoryUrl}";
timeout = 20;
}
];
in {
virtualisation.docker.enable = true;
systemd.tmpfiles.rules = [
"d /var/lib/hermes/.config 0755 hermes hermes -"
"d /var/lib/hermes/.config/tea 0755 hermes hermes -"
"d /var/lib/hermes/.local 0755 hermes hermes -"
"d /var/lib/hermes/.local/share 0755 hermes hermes -"
"d ${aiMemoryDataDir} 0700 hermes hermes -"
"L+ ${aiMemoryDataDir}/config.toml - - - - ${pkgs.writeText "ai-memory-hermes-config.toml" ''
server_url = "${aiMemoryUrl}"
[auth]
bearer_token_file = "${config.age.secrets."ai-memory-hermes-token".path}"
''}"
# Scope marker: Hermes reports its process directory as cwd, so every
# assistant session lands in one project of the `home` workspace
# (coding checkouts below $HOME with their own marker still win).
"L+ /var/lib/hermes/.ai-memory.toml - - - - ${pkgs.writeText "ai-memory-hermes-marker.toml" ''
workspace = "home"
project = "assistant"
[capture]
ignore_paths = ["/run/agenix/**", "**/.env*", "/var/lib/hermes/.hermes/auth*", "/var/lib/hermes/.ssh/**"]
''}"
"L+ /var/lib/hermes/.config/tea/yml - - - - ${pkgs.writeText "tea-yml" ''
logins:
- name: m3ta
@@ -39,7 +72,6 @@ in {
# without runtime pip/uv mutation.
extraDependencyGroups = [
"matrix"
"honcho"
"exa"
"edge-tts"
"voice"
@@ -91,6 +123,11 @@ in {
API_SERVER_PORT = toString (config.m3ta.ports.get "hermes-api");
};
mcpServers.ai-memory = {
command = aiMemory;
args = ["mcp-bridge" "--no-session-id" "--server-url" "${aiMemoryUrl}/mcp" "--data-dir" aiMemoryDataDir];
};
# ── Container mode (podman) ──────────────────────────────────────────
container = {
enable = false;
@@ -217,9 +254,23 @@ in {
silence_duration = 3.0;
};
# ── ai-memory hooks ────────────────────────────────────────────────
# From `ai-memory install-hooks --agent hermes` (3.2.2). ai-memory is
# the one shared memory for every agent; Honcho was dropped in favour
# of it (2026-10-25), only Hermes' small built-in notes remain.
hooks_auto_accept = true;
hooks = {
pre_tool_call = aiMemoryHook "pre-tool-use";
post_tool_call = aiMemoryHook "post-tool-use";
on_session_start = aiMemoryHook "session-start";
pre_llm_call = aiMemoryHook "user-prompt-submit";
on_session_end = aiMemoryHook "stop";
on_session_finalize = aiMemoryHook "session-end";
};
# ── Memory ─────────────────────────────────────────────────────────
memory = {
provider = "honcho";
provider = "builtin";
memory_enabled = true;
user_profile_enabled = true;
memory_char_limit = 2200;
@@ -287,6 +338,26 @@ in {
];
};
# ── ai-memory skills ──────────────────────────────────────────────────
# Hermes ignores MCP server instructions; the routing rules reach the
# model through these skills in ~/.hermes/skills, refreshed on each boot
# so they track the installed ai-memory version.
systemd.services.hermes-ai-memory-skills = {
description = "Install ai-memory skills for Hermes";
before = ["hermes-agent.service"];
wantedBy = ["hermes-agent.service"];
environment = {
HOME = "/var/lib/hermes";
HERMES_HOME = hermesHome;
};
serviceConfig = {
Type = "oneshot";
User = "hermes";
Group = "hermes";
ExecStart = "${aiMemory} --data-dir ${aiMemoryDataDir} install-skills --scope global --agent hermes";
};
};
# ── Activation hardening ─────────────────────────────────────────────
# Upstream (hermes-agent v2026.9.24) deep-merges the Nix settings into
# the runtime-writable config.yaml via a bare `yaml.safe_load()` with no