Skip to content

v0.8.1

Choose a tag to compare

View all tags
@github-actions github-actions released this 29 Jun 18:41
· 2785 commits to main since this release
v0.8.1
This tag was signed with the committer’s verified signature.
alphaonedev AlphaOne
SSH Key Fingerprint: tkfDATcb8+hjhJeI3LvDbwAaks/8QqBMRr6oOWz6iBA
Verified
Learn about vigilant mode.
2a1794a
This commit was signed with the committer’s verified signature.
alphaonedev AlphaOne
SSH Key Fingerprint: tkfDATcb8+hjhJeI3LvDbwAaks/8QqBMRr6oOWz6iBA
Verified
Learn about vigilant mode.

ai-memory v0.8.1 — hardened-patch

The v0.8.0 coordination substrate, review-hardened. A defect-closure patch that takes "shipped" to "shipped, adversarially reviewed, and proven on a second backend": every v0.8.0 development gap closed, a full multi-lane security review with every confirmed finding fixed, and the PostgreSQL + Apache AGE + pgvector path verified live. Self-hosted. MCP-native. Drop-in for v0.8.0.

Why v0.8.1 matters (read this first)

v0.8.0 made the substrate multi-agent. v0.8.1 makes it trustworthy to run — by closing the gaps a careful operator (or auditor) would find first.

This patch is the result of pointing the substrate's own adversarial-review discipline at itself and fixing everything that came back:

  1. Secrets don't leak in, forgotten data doesn't leak out. A credential pasted into a memory is screened on the write path (refuse/redact, configurable); a forget now performs a real erasure fan-out — DLQ cleartext, the transcript-dedup oracle, the vector index, and a signed tombstone so a federation peer can't resurrect it — while an operator's authorized un-forget still round-trips.
  2. The substrate tells the truth about durability. A durable-but-under-replicated write returns 202 Accepted + the replication state, not a misleading 503.
  3. A real security review, all findings fixed. A 7-lane × find→adversarial-verify→triage review surfaced 9 confirmed issues across federation auth, secret-handling, injection/DoS, governance/erasure, and CI supply-chain — all 9 fixed, tested, and closed, each contested call decided by a deterministic 5-agent vote.
  4. Proven on PostgreSQL + Apache AGE + pgvector. The postgres+AGE backend was stood up live and put through 3 green rounds + an AI-NHI dogfood — store, pgvector semantic recall, AGE graph projection, forget/erasure, and the secret-screen, all on real infra.

Same identical API across embedded SQLite and PostgreSQL + Apache AGE, across desktop, server, and on-device.

TL;DR by audience

👤 A. If you just want your AI to remember things (non-technical)

A maintenance release — nothing to relearn. Upgrade and your AI's memory gets safer by default: it won't quietly store a password you paste, and "forget this" now really means gone. Your data still stays on your machine.

brew upgrade ai-memory && ai-memory doctor

Why care: the boring-but-critical safety stuff is now handled for you, without changing how you use it.

🏢 B. If you decide whether to adopt it (C-level / decision-maker)

v0.8.1 is the "we reviewed it and fixed what we found" release — the evidence layer under a production decision.

  • Independently hardened. A multi-lane adversarial security review found 9 issues; 100% are fixed and closed, with the audit trail (issue → fix commit → regression test) public.
  • Right-to-be-forgotten, for real. forget is a genuine erasure across every derived store, with a signed tombstone that blocks peer-driven resurrection — and an operator can still restore intentionally. That's a defensible data-deletion story.
  • Credential hygiene by default. Pasted secrets are screened at the write boundary before they're ever stored, indexed, federated, or exported.
  • Two backends, proven. Validated live on PostgreSQL + Apache AGE + pgvector as well as SQLite — no single-backend lock-in.

Why care: it turns "looks promising" into "reviewed, fixed, and proven on our database."

🛠️ C. If you build or operate it (software engineer / architect — SME)

v0.8.1 is a defect-closure + security-hardening patch (#1821):

  • Gap closure (v0.8.0 dev gaps): G29 write-path secret screen (AI_MEMORY_SECRET_SCREEN_MODE refuse/redact/off — content, title, tags, and metadata-values with a crypto-field carve-out so signatures/JWTs aren't mangled); G30 erasure fan-out + signed forget_tombstones (schema v71); G12 durability now 202 + replication body.
  • Closed audits: MCP governance now fail-closed on the egress sinks (#1685); postgres L2 transcript rehydration parity (#1693).
  • Security review — 9 fixed: federated-signal authorship binding (#1843), field-complete secret screen (#1844), forensic-transcript redaction (#1845), FTS OR-tree DoS cap (#1846), CGNAT SSRF (#1847), archive-restore tombstone gate (#1848), namespace-less forget governance (#1849), audit tail-truncation anchor (#1850), CI workflow_dispatch injection (#1851).
  • Surface (unchanged from v0.8.0 except schema): schema v71 · 100 MCP tools at --profile full / 7 core · 91 HTTP routes (77 unique paths) · 83/85 CLI subcommands · 9 typed link relations · 27-field Memory. SQLite and PostgreSQL+AGE, identical API.

Why care: the gaps you'd file on day one of an audit are already filed, fixed, tested, and closed.

What's new in v0.8.1

🔑 Write-path secret screen (G29)

AI_MEMORY_SECRET_SCREEN_MODE (refuse default / redact / off) screens caller writes — content, title, tags, and metadata string-values — for credential material (PEM/AKIA/ghp_/sk-/xai-/Bearer/JWT), with a crypto-field carve-out (Ed25519 signatures, pubkeys, attestation JWTs) so legitimate signed writes pass unmangled. Federation/recovery paths redact (never refuse) to preserve convergence.

🧹 Erasure that actually erases (G30)

forget fans out across the non-cascaded derived stores (federation DLQ cleartext, transcript-dedup hash oracle, HNSW vector) and writes a signed tombstone that blocks peer-driven resurrection — while an operator's authorized restore still round-trips (the tombstone gates the federation restore path, not first-party un-forget).

🤝 Honest durability (G12)

A locally-durable write that misses quorum returns 202 Accepted with {quorum_met, acks, needed, durability} — the replication state, not a misleading 503.

🛡️ Security review — 9 findings, all fixed

A 7-lane adversarial review (authn/identity, federation, secrets, injection/DoS, governance/erasure, memory-safety, supply-chain) → 9 confirmed → all fixed, tested, closed (#1843–#1851). Contested calls (signal-auth shape, metadata-screen scope, forget governance, truncation anchor, restore/erasure reconciliation) each resolved by a deterministic 5-agent vote.

🐘 PostgreSQL + Apache AGE + pgvector — verified live

Stood up on real infra and run through 3 green rounds + an AI-NHI dogfood: store, pgvector semantic recall, AGE graph projection (memory_graph), tsvector search, forget/erasure, and the secret-screen — all on PostgreSQL 16 + Apache AGE 1.6 + pgvector 0.6. The do-hive provisioning was fixed to install pgvector + build AGE (#1842).

Schema

v70 → v71. One additive migration — the signed forget_tombstones table (G30). Auto-migrates on first open; archive → restore round-trips losslessly.

Distribution channels

Channel Install
GitHub Release this page — binary tarballs for 5 targets + .deb/.rpm + iOS/Android artifacts + SHA256SUMS
crates.io (Rust) cargo install ai-memory --version 0.8.1
Homebrew tap (macOS/Linux) brew install alphaonedev/tap/ai-memory
Fedora COPR (RHEL/Fedora) sudo dnf copr enable alpha-one-ai/ai-memory && sudo dnf install ai-memory
ghcr.io (Docker) docker pull ghcr.io/alphaonedev/ai-memory:0.8.1
npm (TypeScript SDK) npm install @alphaone/ai-memory
PyPI (Python SDK) pip install ai-memory-mcp

Platforms & artifacts

Platform Artifact
macOS x86_64-apple-darwin + aarch64-apple-darwin (Apple Silicon) tarballs; Homebrew
Linux x86_64-unknown-linux-gnu + aarch64-unknown-linux-gnu tarballs; .deb + .rpm; COPR; Docker
Windows x86_64-pc-windows-msvc tarball
iOS ai-memory-ios.xcframework.tar.gz — device (aarch64-apple-ios) + Simulator (arm64 + x86_64) slices
Android ai-memory-android.tar.gz — jniLibs/ with 4 ABIs

Upgrade & compatibility

  • Drop-in for v0.8.0. brew upgrade ai-memory / cargo install ai-memory --force; the live DB auto-migrates v70 → v71 on first open (archive → restore is lossless).
  • New secure default. AI_MEMORY_SECRET_SCREEN_MODE defaults to refuse — a caller write carrying credential material is rejected. Set redact to store a masked copy, or off to disable. See the v0.8.1 CHANGELOG.
  • Durability status change. A quorum-miss on a durable write now returns 202 (was 503). Clients keying off the status code should treat 202 as accepted-and-locally-durable.
  • SDKs (@alphaone/ai-memory, ai-memory-mcp) are published at 0.8.1.

Verification

  • Source provenance: cut from release/v0.8.1; the v0.8.1 tag is signed.
  • Binary integrity: verify downloaded tarballs against the SHA256SUMS on this release page; Docker images carry build provenance; npm publishes with --provenance.
  • Substrate provenance: the audit chain is a tamper-evident cross-row hash chain — now with an on-host watermark that detects tail-truncation (#1850); reflections, signed-events, and forensic bundles are independently verifiable.

Quality gate

Full CI matrix green on the release commit — Linux/macOS/Windows test legs, the SQLite and PostgreSQL+AGE feature gates, clippy -D pedantic + fmt, per-module coverage, iOS/Android cross-compile, MSRV (Rust 1.96), and Docker build · cargo audit clean · the 3 script gates (no-hardcoded-literals, vendor-neutrality, codegraph) green. Validated by a 7-lane adversarial security review with every one of 9 findings fixed and closed, and a live PostgreSQL + Apache AGE + pgvector deployment passing 3 rounds + an AI-NHI dogfood.

From "AIs that coordinate" to "AIs that coordinate — reviewed, hardened, and proven on a second backend." Self-hosted, provider-agnostic, yours.

Assets 17
Loading