v0.21.0 - Release Diagnostics and Host-Neutral Setup

Audrey 0.21.0 - Release Diagnostics and Host-Neutral Setup

This release moves Audrey closer to 1.0 by tightening first-contact setup, release evidence, and the public README front door.

Highlights

• Added npx audrey doctor for runtime, provider, MCP entrypoint, memory-store, and host config diagnostics.
• Added safe npx audrey install --host <host> --dry-run previews for Codex, Claude Code, and generic MCP hosts.
• Reworked the README around Audrey as a local-first memory control plane for agents.
• Added the Audrey wordmark and feature grid assets to the package.
• Added Memory Preflight and Memory Reflexes from the 0.20 work into the release narrative.
• Hardened doctor automation tests across Node 18/20/22 CI while keeping runtime guidance at Node 20+.

Validation

• Master CI passed after merge: Node 18/20/22 on Ubuntu, Windows Node 20, Docker smoke, Python SDK, CodeQL, and Socket checks.
• Local release gates passed: build, typecheck, benchmark regression gate, audit, doctor smoke, host dry-run smokes, demo, and npm pack dry-run.
• Benchmark gate: Audrey 100.0% score and 58.3 points ahead of Vector Only in the local benchmark harness.

Notes

• Audrey's npm runtime target remains Node.js 20+.
• GitHub Actions currently reports Node 20 action deprecation warnings for upstream actions; these are warnings, not release failures.

v0.17.0

Audrey v0.17.0

Audrey 0.17.0 turns the project into a cleaner, release-ready memory platform for Claude Code and agent sidecar deployments.

Highlights

• Added first-run npx audrey init presets for local-offline, hosted-fast, ci-mock, and sidecar-prod.
• Added a PyPI-ready Python SDK as audrey-memory with sync and async clients.
• Added Docker and Compose deployment paths for REST sidecar operation.
• Hardened lifecycle and recall diagnostics, including better shutdown behavior and partial-failure reporting.
• Added retrieval and memory-operations benchmark tracks plus a regression gate via npm run bench:memory:check.
• Tightened the GitHub landing page, README, and public repo metadata for a clearer first-run experience.

Validation

• npm test
• npm run bench:memory:check
• npm run pack:check
• python -m unittest discover -s python/tests -v
• python -m build --no-isolation python
• node mcp-server/index.js doctor
• node mcp-server/index.js --version
• node mcp-server/index.js --help

Upgrade Notes

• Node.js 20+ is now the supported baseline.
• Docker deployments should start from .env.docker.example.
• The README now focuses on quick start, operations, and packaging instead of long-form benchmark detail.

Full Changelog: v0.16.1...v0.17.0

v0.16.1 — Windows MCP fix

Fix

• npx audrey install now wraps the MCP command in cmd /c on Windows, fixing the spawn error where npx (a cmd script) can't be executed directly without a shell.

Before

{ "command": "npx", "args": ["audrey"] }  // fails on Windows

After

{ "command": "cmd", "args": ["/c", "npx", "audrey"] }  // works everywhere

Re-run npx audrey install to fix your registration.

464 tests, 29 test files, 0 failures.

v0.16.0

Audrey v0.16.0

Version bump for npm publish. All v0.15.0 features included — see v0.15.0 release notes for the full changelog.

Additional fixes since v0.15.0 publish

• Fix circular self-dependency in package.json (audrey depending on itself)
• greeting and reflect CLI subcommands for hook integration
• Shared resolveLLMConfig() — dream/reflect CLI now respects AUDREY_LLM_PROVIDER env var
• LLM JSON parsing strips markdown code fences from any provider
• Better Anthropic API error messages (includes response body)

Install

```bash
npx audrey install
```

463 tests, 29 test files, 0 failures.

v0.15.0 — Production Ready

Audrey v0.15.0 — Production Ready

The biological memory architecture for AI agents is now production-hardened and ready for the world.

Highlights

Dream Cycle — npx audrey dream runs the full biological sleep analog: cluster episodic memories → extract semantic/procedural principles via LLM → apply forgetting curves → report health. Wire it into your session hooks for automatic memory maturation.

CLI Subcommands for hook integration:

• npx audrey greeting — session-start briefing with mood, principles, recent memories
• npx audrey reflect — reads conversation turns from stdin, forms lasting memories, runs dream cycle
• npx audrey dream — consolidation + decay + introspect
• npx audrey status — memory store health at a glance

Production Hardening:

• Consolidation is fully transactional — mid-run failures roll back cleanly
• Recall gracefully degrades per memory type (episodic/semantic/procedural search independently)
• sqlite-vec crash guard when vector tables are empty
• LLM JSON parsing handles markdown code fences from any provider
• Input validation: empty content rejected, 50K char limit, forget requires exactly one target

Dual Memory Routing: Consolidation now routes principles to the correct table — type: 'procedural' goes to procedures with trigger conditions, type: 'semantic' goes to semantics.

SDK Surface: Full library exports for using Audrey as a dependency without the MCP server — all embedding providers, LLM providers, database utilities.

Full-Fidelity Export/Import: Snapshots now preserve consolidation metrics, run metadata, and config. Import uses batch embedding in a single atomic transaction.

Stats

• 463 tests, 29 test files, 0 failures
• 13 MCP tools + 7 CLI subcommands
• Supports local (384d), Gemini (3072d), and OpenAI (1536d) embeddings
• LLM providers: Anthropic, OpenAI, Mock

Install

npx audrey install

Co-built by

• Tyler Eveland (@Evilander)
• Claude Opus 4.6 (Anthropic)
• Codex GPT-5.4 (OpenAI)