Skip to content

v1.0 GA criteria

Publishing the criteria, not the date. Every 0.6.x / 0.7.x / 0.8.x release is step N toward the contract below. When every line in this document is green, and has been green for the qualifying window, the v1.0.0 tag cuts.

This document spans three repositories — ai-memory-mcp (the substrate), ai-memory-ship-gate (substrate validation against live infrastructure), and ai-memory-ai2ai-gate (this repo — multi-agent integration validation) — because v1.0 is a statement about the whole stack, not any single component.

Authority: THE STANDARD governs process; this document is the technical target.

1. Substrate — ai-memory-mcp

1.1 API contract frozen (breaking-change budget = 0)

  • The HTTP API surface (/api/v1/*) is stable: every endpoint documented in handlers.rs carries a deprecation policy instead of a removal policy.
  • The MCP tool surface (23+ tools defined in mcp.rs) is stable: no tool rename, no parameter rename, no response shape change between v1.0 and any v1.x.
  • The CLI surface (26+ subcommands) is stable: same rule.
  • Any additive field is explicitly marked additive in the release notes and defaults to backwards-compatible behavior.
  • The wire format for federation (POST /api/v1/sync/push) is versioned and dual-reads until v2.

1.2 Test posture

  • cargo fmt --check + cargo clippy -- -D warnings -D clippy::all -D clippy::pedantic + cargo test + cargo audit green on ubuntu / macos / windows for three consecutive tagged releases.
  • Integration test count ≥ 200, all green. Baseline today: 183 (see tests/integration.rs).
  • Benchmark suite (cargo bench) runs clean; no regression > 10% on recall, search, detect_contradiction compared to v0.6.0 baseline.
  • No unsafe, no panic!() in non-test code, no .unwrap() on untrusted input.

1.3 Security posture

  • No open cargo audit finding at any severity.
  • Signed commits on every merge to main / release/*. Signed release tags (via jbridger2021 key, registered on crates.io and GitHub).
  • mTLS support (Layer 2 in v0.6.0, carried forward) has a documented production-deployment runbook in docs/.
  • SBOM published alongside every release (SPDX JSON attached to the GitHub release).
  • Secret-scanning enabled on all repos; zero detections carry forward into a tag.
  • No unresolved CVE advisories against runtime dependencies.

1.4 Observability

  • Prometheus /metrics surface covers all five subsystems: writes, recalls, GC, federation (acks + quorum-misses), cache. Standing policy — any new handler that mutates state adds a counter.
  • OTEL_* / OpenTelemetry export behind a feature flag, with a documented Grafana dashboard and Loki query set.
  • Every error path returns a classified ApiError (not a raw string). Error codes are documented.

1.5 SDK coverage

  • Python SDK — 1.0-pinned, published to PyPI with a semver guarantee. Full HTTP + MCP coverage. Tested against three consecutive ai-memory-mcp releases.
  • TypeScript SDK — 1.0-pinned, published to npm, same guarantee.
  • Rust crate — ai-memory-client sibling crate separate from the binary, published to crates.io.
  • Each SDK has a getting-started tutorial + ≥ 5 runnable cookbook examples.

1.6 Documentation

  • Full API reference (mkdocs or similar) covers every HTTP endpoint + every MCP tool + every CLI subcommand.
  • Operations guide: deployment, scaling, backup, restore, migration, upgrade-path, disaster recovery.
  • Security hardening guide: TLS, mTLS, SQLCipher, secret rotation, agent identity.
  • Federation guide: topology choice, quorum tuning, partition recovery, cross-version compatibility matrix.

2. Substrate validation — ai-memory-ship-gate

  • Phases 1 – 4 green on three consecutive releases:
  • Phase 1: single-node smoke (write / recall / GC / restart).
  • Phase 2: 3-node federation (quorum writes, catchup, partition convergence).
  • Phase 3: SAL cross-backend (SQLite ↔ Postgres migration, 1000-memory round-trip, zero errors).
  • Phase 4: chaos harness (SIGSTOP, network blip, disk-full, kill -9).
  • Evidence published live under alphaonedev.github.io/ai-memory-ship-gate/. Every phase artefact is a JSON file peer-reviewable out-of-band.
  • Ship-gate dispatch is triggered on every tag; failure blocks GHCR / crates.io / Homebrew / PPA publication.

3. Multi-agent integration — ai-memory-ai2ai-gate (this repo)

3.1 Full-matrix certification

Homogeneous-framework matrix:

off tls mtls
ironclaw 34 / 34 × 3 35 / 35 × 3 36 / 36 × 3
hermes 34 / 34 × 3 35 / 35 × 3 36 / 36 × 3

Mixed-framework matrix (post-topology work, gated by this repo's terraform changes):

off tls mtls
mixed (ironclaw + hermes) 34 / 34 × 3 35 / 35 × 3 36 / 36 × 3

"N / N × 3" = three consecutive overall_pass = true runs at full scenario coverage. See the certification threshold on the home page for the counting rule — a single red resets the streak for that cell.

3.2 Framework expansion

  • Phase 2 — Claude Managed Agents (per roadmap.md): claude-managed agent group green on Claude Haiku 4.5 default, across all three transport modes, for ≥ 1 full matrix sweep. Adds an independent runtime to the "framework-agnostic" claim.
  • Phase 3 — cross-cloud (stretch, may land in v1.1): one non-DigitalOcean provider (AWS, GCP, Hetzner) proven as a drop-in target.

3.3 Adversarial + Byzantine coverage

  • Suite G scenarios (S20 – S27) green across the matrix. mTLS happy-path, anonymous rejected, identity spoofing, malicious content fuzz, Byzantine peer, clock skew — all pinned.
  • Zero known bypasses of the scope visibility matrix under the scope-enforcement contract (Task 1.5).

3.4 Observability surface

  • Every failing scenario produces a one-line failure narrative in analysis/run-insights.json. No silent "0-byte report" scenarios (see historical bug ai-memory-ai2ai-gate#2).
  • baseline_pass / overall_pass emitted per-campaign-per-cell, aggregated into a public dashboard.

4. Release discipline

  • Every release from v0.6.0 forward has landed under the AlphaOne Root Cause Analysis Standard v1 (authored 2026-04-21). v1.0 cuts only after ≥ 10 consecutive releases under the standard without a post-release RCA.
  • Every release carries an AI involvement section (per AI_DEVELOPER_WORKFLOW.md §8.2) disclosing the AI NHI(s) involved, the authority class exercised, and the biologic review path.
  • Every commit is signed, every PR has a Co-Authored-By trailer, every merged PR has green CI across all three platforms.

5. Explicit non-goals for v1.0

To keep the contract honest, these are deliberately out of scope for v1.0:

  • Multi-tenant SaaS. The HTTP API is multi-tenant per request (agent_id) but not per tenant boundary; that's a v1.x item tracked as ai-memory-mcp#148 and the upcoming attested-agent-id work.
  • Cross-cloud federation as a first-class product. v1.0 validates on one cloud (DigitalOcean) end-to-end; cross-cloud is a stretch (see §3.2).
  • Human-in-the-loop agent supervision UX. Governance primitives ship (pending / approve / reject, policy rules). A polished supervisor UI does not; that is AgenticMem commercial surface.
  • LLM-native query planner. The smart/autonomous tiers use LLMs opportunistically (auto-tag, contradiction detection, query expansion) but the core recall pipeline remains deterministic in v1.0.
  • Formal verification of the CRDT / LWW semantics. v1.0 ships vector-clock + LWW with documented edge-cases; formal proofs are a v1.x research item.

6. How we keep this honest

This document lives in version control and is re-rendered on every merge to main. Any of the following changes the contract and requires a corresponding PR that updates this page, the home page's certification threshold section, and the affected release notes:

  • Changing the scenario corpus size (new suite, removed scenario).
  • Adding or removing a transport mode.
  • Adding or removing an agent framework to the matrix.
  • Changing the streak length for certification (currently 3).
  • Changing the breaking-change budget (currently 0).

If you find this document out of date relative to what's actually shipping, that's a bug — file an issue against the docs label.