Benchmark: Mutation Usability Regression Campaign (2026-09-08)

MemAuthority Benchmark — Mutation Usability Regression

Complete audit of cross-session handoff reliability with 10 rounds of testing.

Results

Metric	Result
Mutation operations	426 total
Mutation errors	2 (0.5%)
Total runs	30/30 PASS
Total sessions	90/90 PASS
Human semantic review	30/30 confirmed
Independent audit	Claude Opus 5, HIGH confidence

Agent Stacks (10 rounds each)

• Codex CLI + GPT-5.6 Sol: 10/10 runs, 0/136 mutation errors (0.0%)
• Codex CLI + GPT-5.6 Luna: 10/10 runs, 0/138 mutation errors (0.0%)
• Pi Agent + Gemini 3.8 Flash: 10/10 runs, 2/152 mutation errors (1.3%, both recovered)

Key Finding

All tested Agent Stacks successfully preserved and recovered project state across fully independent sessions. The corrected mutation error rate of 0.5% (2/426) demonstrates high API usability.

Audit Correction

Original report: 18/426 mutation errors (4.2%)
Audited actual: 2/426 mutation errors (0.5%)

Comprehensive MCP log analysis with explicit criteria corrected the count. Both errors occurred in the same run, same tool, same validation issue. Agent recovered successfully. See audit documentation for complete methodology.

Documentation

Complete audit documentation is available in the repository:

• Audit Summary
• Complete Audit Report
• Semantic Review Confirmation
• Chinese Summary
• Full Documentation

Evidence Download

This release includes complete evidence artifacts for all 30 runs:

• 📦 benchmark-v1-mutation-usability-regression-runs-20260908.tar.gz (326 MB)
  • 30 complete run directories
  • Full MCP request/response logs
  • Project and vault snapshots per stage
  • Frozen inputs and configurations
  • Automatic acceptance results
  • Semantic review confirmations

Product Information

• Product: MemAuthority (post-1.3.2 with mutation usability fixes)
• Commit: 94eb36a
• Binary SHA-256: 6981dfdc57c37a55da5c8fe6b0277f4739fe52780a3c1a19e887c428432c555e

Audit Certification

• Auditor: Claude Code (Opus 5) / Claude (Anthropic)
• Model: claude-opus-5[1M]
• Audit Date: 2026-09-08
• Coverage: 100% (30 runs, 90 sessions, 426 operations)
• Methodology: Comprehensive MCP log analysis + semantic verification
• Confidence: HIGH
• Determination: PASS
• Recommendation: APPROVED FOR PUBLICATION

---

Citation:

MemAuthority Benchmark v1 — Mutation Usability Regression Campaign
Case: handoff-pending-01, 10 rounds, 3 Agent Stacks
Executed: 2026-09-08
Audited: 2026-09-08 by Claude Code (Opus 5)
Result: 30/30 PASS, 426 mutation operations, 2 errors (0.5%)
Product: MemAuthority commit 94eb36aa (post-1.3.2 with usability fixes)

MemAuthority v1.3.2

MemAuthority v1.3.2 is a compatibility-preserving packaging release that prepares the project for the Official MCP Registry.

What changed

• Added six native MCP Bundle (MCPB) packages for Windows, Linux, and macOS on x64/ARM64.
• Added validated Official MCP Registry metadata under server.json with identity io.github.iasi777/memauthority.
• MCPB installs remain read-only by default; write access must be enabled explicitly.
• Added reproducible packaging: repeated builds from the reviewed source baseline produce the same MCPB SHA-256 values.
• Added public GitHub Actions verification for MCPB packaging and Registry metadata; it does not publish automatically.

Compatibility

There is no intended Vault format, MCP tool/resource, CLI command, transport/auth, or v1.x Go module compatibility change beyond the patch version identity.

The v1.x Go module identity remains github.com/iasi777/v-memory for compatibility.

Install with Go

go install github.com/iasi777/v-memory/cmd/memauthority@v1.3.2

MCP Registry packaging

The attached .mcpb files are the exact artifacts referenced by the checked-in server.json. SHA256SUMS contains the release hashes.

Contract identity:

7ee8e2b1af9ba7bde7843ea04bb1df97f2ca7ef5bd2f4f777cac4bac517d2964

MemAuthority v1.3.1

MemAuthority v1.3.1 is the current stable release of the Git-backed, conflict-safe long-term memory authority for AI agents via MCP.

Highlights:

• memauthority is the primary CLI, with v-memory retained as a v1 compatibility command.
• MCP implementation identity is MemAuthority / 1.3.1.
• Git-backed Authority preserves auditable history while CAS, idempotency, transaction journaling, and recovery protect memory updates from stale writes, retries, and interruptions.
• Runtime metadata remains optional and progressive: ordinary single-machine setups keep the smaller default surface, while existing runtime-aware Vaults remain compatible.
• Preferred MEMAUTHORITY_* environment variables coexist with legacy VM_* aliases.
• Linux, macOS, Windows, and Linux/ARM64 build paths are validated in CI.
• Frozen-contract manifest integrity and runtime MCP snapshot checks protect the public compatibility surface from accidental drift.

Install the primary CLI:

go install github.com/iasi777/v-memory/cmd/memauthority@v1.3.1

Existing automation may continue using the compatibility CLI:

go install github.com/iasi777/v-memory/cmd/v-memory@v1.3.1

The v1.x Go module identity intentionally remains github.com/iasi777/v-memory for compatibility even though the canonical repository and product name are now MemAuthority.

Maintenance note:

• v1.3.1 removes maintainer-specific production deployment workflow metadata from the public source/module distribution. No credential values were exposed, and this change does not alter the product runtime, Vault format, MCP tool/resource schema, CLI command surface, or v1.x module compatibility.
• v1.3.0 was already published and cached by the Go module proxy before that cleanup; it remains an immutable historical archive and is marked as superseded by v1.3.1.

Release identity:

• source baseline: 532034f4d0dbbe5f59437f8ab1a1015b195b4eb0
• freeze commit: 3bb4933685e138839ff1c0279be3fb3f0ab83de7
• contract identity: 79036ecac337e51cff679a09477bdbdba248c975d11c613e1892dd002092e94c
• baseline CI: 33289714731
• freeze CI: 33289841521

The current frozen public contract is docs/contract/v1.3.1/; earlier released contract snapshots remain immutable.

MemAuthority v1.3.0

Superseded by v1.3.1. The v1.3.0 source archive included maintainer-specific production deployment workflow metadata (no credential values). v1.3.1 removes that private deployment plane from the public repository/module distribution. New installations should use v1.3.1.

MemAuthority v1.3.0 is the first release under the MemAuthority brand, renamed from V-Memory to establish a unique, directly searchable project identity while preserving the v1 compatibility line.

Highlights:

• Canonical repository renamed to iasi777/memauthority; the former GitHub repository URL redirects to the new location.
• New primary executable: memauthority.
• Legacy v-memory executable remains supported with the same v1 command surface.
• The v1.x Go module identity intentionally remains github.com/iasi777/v-memory, preserving existing module/install semantics through the repository redirect.
• Preferred MEMAUTHORITY_* environment variables were added; legacy VM_* names remain accepted, with the new name taking precedence when both are set.
• MCP implementation identity is now MemAuthority / 1.3.0; tool names, Vault semantics, transaction safety, runtime behavior and structured errors remain compatible with v1.2.
• Durable historical identifiers such as v-memory: mechanical commit prefixes, V-Memory MCP <v-memory@v-memory.invalid>, and <!-- v-memory-entry:... --> remain unchanged by design.
• Added a permanent CI test that compares the runtime-enabled MCP tool/resource surface against the frozen v1.3 JSON contract snapshots.

Install the primary CLI:

go install github.com/iasi777/v-memory/cmd/memauthority@v1.3.0

Existing automation may continue using:

go install github.com/iasi777/v-memory/cmd/v-memory@v1.3.0

Release identity:

• source baseline: a9598d57f079879bf55693f5c68eb189b7615cf3
• freeze commit: 747d40eb18a5adb431581ec3a12f0eb20114b725
• contract identity: b0c2fc501933445f00c5ac3ebfd31e5773a5d6ecf750c967d28fd1e5d2e2fafb
• baseline CI: 33289239698
• freeze CI: 33289403893

The historical v1.1 and v1.2 V-Memory contracts remain unchanged under docs/contract/v1.1/ and docs/contract/v1.2/. The current frozen public contract is docs/contract/v1.3/.

This release does not deploy production or rename the existing production OAuth/deployment domain and gateway protocol identifiers.

V-Memory v1.2.0

V-Memory v1.2.0 is the first cross-platform validated minor release after the public v1.1 baseline.

Highlights:

• Native CI validation on Linux, macOS, and Windows, including Windows durability and Git line-ending compatibility fixes.
• Managed revisions now bind to Git Authority bytes rather than checkout-specific line-ending transformations.
• Optional runtime metadata is progressive: runtime-free Vaults expose the smaller default MCP surface; existing Vaults with runtime metadata remain auto-enabled; new advanced setups can opt in with --runtime-enabled.
• host_id is no longer tied to maintainer machines and is now optional/user-defined.
• Single-environment runtime input may omit environment_id; it canonicalizes to default. Multi-environment manifests still require explicit unique IDs.
• Onboarding and migration docs now recommend leaving runtime metadata disabled unless durable work/deployment topology is genuinely useful.

Release identity:

• source baseline: 0b46dc0a15cf8c7981567586b5f8f3d0caf996bf
• freeze commit: af6c6fad7d4b03be96db041c58251d6ebe25bd51
• contract identity: eb4f28e75ee7231c4b7882a8e93adf20b2eda05b41f93978fcb9a3ae7007c566
• baseline CI: 33166991263
• freeze CI: 33167238099

The frozen v1.1 contract remains unchanged under docs/contract/v1.1/; the current public contract is docs/contract/v1.2/.

V-Memory v1.1.0

V-Memory v1.1.0 is the first public source release in the rebuilt repository.

Highlights:

• Context-first, recall-on-demand MCP capability guidance with no fixed route -> search -> read ritual.
• Human-oriented titles and point-of-use contracts for all 16 MCP tools.
• Dependency-aware aggregated runtime validation with additive validation_errors[].
• Fresh static runtime validation before INDEX/CAS reads.
• Handoff checklist mutation through stable item_ref values.
• Versioned public contract under docs/contract/v1.1/.
• English-primary, Chinese-paired user documentation and a runnable sample Vault.

Release identity:

• Public source baseline: 4bc987e60ce926d27f6fe28e6dc05133a6ed8be4
• Freeze/tag commit: b6e2455d1ca8a6f9cbe669adbc1979aae33efa55
• Contract identity SHA-256: d84a43223df001bd3f7b846192344c188d33962aaf2d4a1fbbaa20b7f76314bb
• Freeze CI: run 33159882842 (success)

Earlier private development, deployment, and freeze records are intentionally outside this public Git identity. Production deployment provenance remains private operational state and does not define the public source release.