v0.1.1

OCM v0.1.1 — DRAFT release notes

Status: DRAFT — written 2026-06-11. Paste into the GitHub draft release when
tagging v0.1.1. Tagging requires the version bump in Cargo.toml /
frontend/package.json / tauri.conf.json per docs/release-process.md.

Headline: the memory loop is now measured, not promised

v0.1.0 shipped the architecture; v0.1.1 ships the first evidence. The bench
suite went from scaffold to working instrument, and the central claim — that
library-driven memory retrieval makes a small local model materially better —
now has a CONFIRMED verdict behind it.

Benchmarks (the headline numbers)

• amnesia-ab (NEW, first memory sandbox to run) — CONFIRMED.
  Memory ON vs OFF on llama3 8B Q4 + mxbai-embed-large, 62-memory corpus with
  cross-project distractors, 20 tasks, objective key-fact scoring:
  94.2% fact recall with memory vs 2.5% without; retrieval hit rate
  100%. The OFF arm's failure mode is confident fabrication, not ignorance —
  the memory loop is the difference between correct specifics and plausible
  lies on an 8B model.
• aider-repomap-fidelity — ACTIVE, 59.2% CONFIRMED (#53).
• sandbox-e schema compression and sandbox-i Mem0 encryption overhead
  flipped ACTIVE (#51, #52); frontier-comparison sandboxes E–I landed as
  INACTIVE hypothesis stubs (#48).
• run-all orchestrator + cross-sandbox comparison table (#47) and the
  comprehensive bench guide docs/benchmarking.md (#49).

Inference

• NEW: Ollama backend adapter (ocm-inference::ollama). Bridges OCM to an
  existing Ollama daemon via its native NDJSON API — the zero-extra-process
  path for the largest installed base of local-model users. Parser tests are
  pinned to verbatim captures from a live daemon. Explicit construction for
  now; daemon settings wiring (backend = "ollama") is the follow-up.

Models

• Registry is now 3 models, 3/3 SHA256-verified — the unhashed Qwen3
  entries were dropped (#50) and return when verified hashes are available.
  The "downloads no-op on empty hashes" pre-release blocker is gone.

Docs

• README corrected to match the shipping registry (was still claiming 5 GGUFs
  and an open hash blocker).

Known limitations (unchanged from v0.1.0)

• Unsigned installers (codesigning lands in v4) — expect OS warnings.
• Daemon expects external inference + Mem0 processes; supervised spawn is the
  next UX milestone.
• Mesh transport remains a v2 trait scaffold.

v0.1.0

OpenCircuitModel — v1 release artifacts. Unsigned in v1; codesigning lands in v4.

Install

• macOS: download the .dmg, drag OCM to Applications. First launch will warn about an unidentified developer; right-click → Open to bypass once.
• Windows: download the .msi, run installer. SmartScreen may warn; click "More info" → "Run anyway".
• Linux: download the .deb (Debian/Ubuntu) or .AppImage (any distro). sudo dpkg -i ocm_*.deb or chmod +x the AppImage.

After install

• The daemon listens on http://127.0.0.1:7300 (configurable in Settings)
• Point any OpenAI-compatible client (Cline, Continue.dev, ChatGPT desktop) at the localhost URL
• For chat memory, install Mem0's OpenMemory MCP locally (optional but recommended)

What's Changed

• Phase 1: Foundation — Tauri shell, paths, settings, CI by @OpenCircuitDev in #1
• OCM v1 daemon stack: phases 2-8 (cascade recovery, conflicts resolved) by @OpenCircuitDev in #17
• Phase 0: Bench framework scaffold + first isolation sandbox by @OpenCircuitDev in #2
• VM/cloud deployment policy (spec v0.5 row 31) + multipass-fleet sandbox stub by @OpenCircuitDev in #10
• docs(readme): reflect v1-mostly-built state by @OpenCircuitDev in #15
• ci: path-filter Rust workflow to crates/Cargo only by @OpenCircuitDev in #18
• docs: add CONTRIBUTING.md by @OpenCircuitDev in #19
• feat(bench): 7 sandbox stubs from queued research hypotheses by @OpenCircuitDev in #20
• docs+chore: spec v0.6 hygiene + registry-hash population script by @OpenCircuitDev in #21
• chore: add scripts/smoke-test-daemon.sh — v1 verification harness by @OpenCircuitDev in #22
• feat(bench): split ACTIVE/INACTIVE in dry-run-all + add list-inactive command by @OpenCircuitDev in #23
• chore: .editorconfig + .gitattributes (cross-platform consistency) by @OpenCircuitDev in #24
• docs: add CODE_OF_CONDUCT.md by @OpenCircuitDev in #25
• docs: GitHub issue + PR templates by @OpenCircuitDev in #26
• chore: Dependabot config + git-blame-ignore-revs by @OpenCircuitDev in #27
• chore(models): populate SHA256 hashes for 3 of 5 v1 registry entries + Python hash script by @OpenCircuitDev in #42
• fix(release): install tauri-cli before cargo tauri build by @OpenCircuitDev in #43
• fix(release): drop beforeBuildCommand (CI does it explicitly) by @OpenCircuitDev in #44
• fix(release): correct artifact paths (workspace-root + multi-line glob) by @OpenCircuitDev in #45
• fix(release): drop macos-13 from matrix (Intel runner unavailable) by @OpenCircuitDev in #46

Full Changelog: https://github.com/OpenCircuitDev/opencircuitmodel/commits/v0.1.0