Infrastructure inventory¶
This page enumerates every component version pinned or in active use across the a2a-gate harness, the ai-memory substrate, and the agent frameworks exercised per campaign. Kept up to date when any version pin moves; last refreshed 2026-04-21.
Test substrate — ai-memory¶
| Component | Version | Source of truth |
|---|---|---|
| ai-memory (latest tagged) | v0.6.1 |
GitHub release |
| ai-memory (under test in the a2a gate) | v0.6.1 until tagged v0.6.2 |
workflow ai_memory_git_ref input |
| Rust MSRV (source build) | 1.93 |
Cargo.toml rust-version |
Rust in CI (x-platform cargo clippy + test) |
stable via dtolnay/rust-toolchain@stable |
ci.yml |
| Rust pin for MSRV drift gate | 1.93 (Ubuntu 26.04 rust-defaults) |
ci.yml msrv job |
| Rust edition | 2024 |
Cargo.toml |
ai-memory's stable-Rust CI uses whatever stable points to on the
GitHub runner image at CI time (today 1.94.x). The MSRV CI
job separately pins rustc 1.93 and runs cargo check --locked
so any accidental use of post-1.93 APIs fails CI before a tag-time
PPA upload attempts it.
Agent frameworks (per campaign)¶
| Framework | Role | Version / ref | Install path |
|---|---|---|---|
| IronClaw | primary Rust agent (replaces OpenClaw 2026-04-21) | latest release from nearai/ironclaw (0.26.0 as of this writing) | direct tarball download ironclaw-x86_64-unknown-linux-gnu.tar.gz into /usr/local/bin/ironclaw (bypasses NEAR AI's installer script due to platform-detection bug) |
| Hermes | Python agent | pinned ref HERMES_INSTALL_REF=main from NousResearch/hermes-agent |
curl install.sh --skip-setup + pip install python-dotenv==… pinned patch |
| OpenClaw | LEGACY — retained for historical dispatch reproduction | openclaw/openclaw via install.sh --install-method git |
scripts/setup_node.sh legacy branch |
All three are agent-agnostic from the a2a-gate's perspective — the scenarios drive ai-memory's HTTP surface directly, so the framework under test only affects baseline attestation + substrate provenance, not scenario logic.
LLM backend¶
| Model SKU | Provider | Base URL | Default reasoning? |
|---|---|---|---|
grok-4-0709 (= Grok 4.2 reasoning) |
xAI | https://api.x.ai/v1 |
yes |
grok-4-fast-non-reasoning |
xAI | same | no (override for cost-optimized smoke runs only) |
All agent groups use the same SKU per campaign so A2A
comparisons hold. Configured via A2A_GATE_LLM_MODEL workflow
input; default grok-4-0709.
DigitalOcean infrastructure (per 4-node campaign)¶
| Resource | Pin | Notes |
|---|---|---|
| Droplet base image | ubuntu-24-04-x64 |
DO's Ubuntu 24.04 LTS ("noble") image |
| Agent droplet size — ironclaw | s-2vcpu-4gb |
Basic tier, 2 vCPU, 4 GB RAM, ~$0.03/hr — proven on the 3-green r25/26/27 streak |
| Agent droplet size — hermes | s-4vcpu-8gb (hard-coded; see #40) |
Basic tier, 4 vCPU, 8 GB RAM, ~$0.06/hr — Python LLM stack needs 8GB headroom; agent_droplet_size workflow input is ignored for agent_group=hermes |
| Memory node droplet size | s-2vcpu-4gb (hard-coded in terraform/main.tf) |
matches ironclaw agents — no upsell needed |
| Region | nyc3 (default; overridable via workflow input) |
|
| VPC CIDR — ironclaw campaigns | 10.251.0.0/24 |
primary Rust agent (replaces openclaw) |
| VPC CIDR — hermes campaigns | 10.252.0.0/24 |
Python agent |
| VPC CIDR — openclaw campaigns (legacy) | 10.253.0.0/24 |
retired-default |
| Ship-gate VPC CIDR | 10.250.0.0/24 |
ai-memory-ship-gate reserves this — we deliberately avoid it |
Per-droplet software stack¶
Provisioned per scripts/setup_node.sh:
| Component | Version | Source |
|---|---|---|
| OS | Ubuntu 24.04 LTS (noble) | DO base image |
ai-memory |
${AI_MEMORY_GIT_REF} (workflow input) |
GitHub Releases binary tarball (ai-memory-x86_64-unknown-linux-gnu.tar.gz) |
| PostgreSQL (ironclaw only) | 15 (target) via PGDG apt repo |
apt.postgresql.org.sh adds the repo, then apt-get install postgresql-15 postgresql-15-pgvector postgresql-contrib-15 |
| pgvector (ironclaw only) | postgresql-15-pgvector (matches PG major version) |
PGDG apt repo (not default noble apt) |
| Node.js / npm (openclaw only, legacy) | whatever apt install nodejs npm pulls |
used by openclaw's install.sh |
| Python 3 (hermes only) | system python3 (3.12 on noble) | used by hermes-agent |
python-dotenv (hermes only) |
pinned patch version in setup_node.sh | upstream hermes install doesn't install it but needs it at runtime |
systemd, curl, jq, sqlite3 |
Ubuntu default | |
| UFW | DISABLED + verified disabled | ship-gate lesson; baseline invariant |
| iptables | flushed to ACCEPT INPUT/OUTPUT/FORWARD | baseline invariant |
| Dead-man switch | shutdown -P +480 (8-hour auto-halt) |
cost cap |
ai-memory federation wiring¶
Every node runs ai-memory serve with:
--host 0.0.0.0 --port 9077--quorum-writes 2(W=2 of N=4 mesh)--quorum-peers= comma-separated URLs of the other 3 nodes--db /var/lib/ai-memory/a2a.db
Default feature tier: semantic (HuggingFace-Hub embedding model
auto-downloaded at first boot).
GitHub Actions CI¶
| Runner | Purpose | Image |
|---|---|---|
ubuntu-latest |
ai-memory Check x86_64, release builds |
currently ubuntu-24.04 |
ubuntu-24.04-arm |
ai-memory cross-compile for aarch64-unknown-linux-gnu |
GitHub hosted ARM |
macos-latest |
ai-memory Check + release for aarch64-apple-darwin / x86_64-apple-darwin |
currently macos-14 |
windows-latest |
ai-memory Check + release for x86_64-pc-windows-msvc |
| Tool | Version |
|---|---|
| Terraform | 1.9.5 (pinned via hashicorp/setup-terraform@v3) |
actions/checkout |
v5 |
dtolnay/rust-toolchain |
@stable (for check matrix) / @master + explicit 1.93 (for MSRV job) |
Swatinem/rust-cache |
v2 |
cargo-audit |
installed fresh per run via cargo install cargo-audit --locked |
Distribution targets¶
| Channel | Compiled where? | Target audience MSRV-sensitive? |
|---|---|---|
| GitHub Release tarballs | ai-memory CI (Rust stable = 1.94 today) | no — binary |
| Fedora COPR | ai-memory CI (repackages the x86_64 / aarch64 Linux tarballs) | no — binary |
| Homebrew | ai-memory CI | no — binary |
| Docker GHCR | ai-memory CI | no — binary |
| Ubuntu PPA | Launchpad's build farm (rustc from rust-defaults package) |
yes — Rust 1.93 required |
cargo install ai-memory from crates.io |
user's machine | yes — Rust ≥ 1.93 |
Ubuntu PPA target release¶
- Series:
resolute(Ubuntu 26.04 LTS — Resolute Raccoon). - 26.04 ships
rust-defaultspointing at rustc 1.93 (with 1.91 + 1.92 also in the archive for side-by-side coexistence). - The PPA CI step rewrites
debian/changelogat tag time to set the series, thendebuild -S+dputuploads the signed source package. Launchpad asynchronously compiles. - Earlier releases (noble 24.04 et al.) are NOT supported by new PPA uploads as of 2026-04-21. Users on those releases must upgrade to 26.04 to receive new .debs.
ai-memory runtime requirements¶
For end users installing from any path:
| Requirement | Version |
|---|---|
| OS | Linux (glibc ≥ 2.38), macOS 13+, Windows 10+ |
| RAM (keyword-only deployment) | 100 MB typical |
| RAM (semantic tier with embedder loaded) | 300–500 MB typical |
| Disk | 200 MB for binary + ~/.ai-memory DB (grows with corpus) |
| Outbound network | api.x.ai (for xAI Grok tests), HuggingFace Hub (one-time model download), federation peer URLs |
Pin-bump procedure¶
When any version in this table changes:
- Update the corresponding code pin (Cargo.toml / setup_node.sh / ci.yml / terraform/main.tf).
- Add a
CHANGELOG.mdentry in the originating repo. - Update the table row in this doc.
- If the change materially affects the baseline claim (e.g. MSRV bump), re-run a full a2a-gate campaign in each group (ironclaw + hermes) before tagging the next release.
Authoritative version list (machine-readable)¶
For automation, campaign.meta.json emitted by every campaign carries
the resolved version pins for THAT run:
{
"ai_memory_git_ref": "v0.6.1",
"agent_group": "ironclaw",
"infra": {
"provider": "digitalocean",
"region": "nyc3",
"droplet_size": "s-2vcpu-4gb",
"topology": "4-node federation mesh (W=2/N=4)"
},
"llm_model": "grok-4-0709"
}
Run-specific attestation also lives per-node in a2a-baseline.json: