What changed, and when.

Every released version of NeuralVault, newest first. This is the same list the release emails are built from, so it says exactly what those said. What is coming next is on the roadmap.

1.0.0-betacurrent release
20 August 2026released
54changes recorded
2major releases

Editions and download · What is coming next

1.1the Personal releasenot released yettarget: early to mid-September

What the next release adds, in short. See how far along each piece is.

  • A rebuilt interface
  • A signed installer
  • A front door
  • Signing in replaces licence files
  • Account deletion, and telemetry that is off
  • Your access follows your subscription
  • Memory and skills released as Fair SourceFree
  • NeuralVault PersonalPersonal
  • Included credits, and usage you can seePersonal
  • NeuralVault routing in the model pickerPersonal
  • Reasoning effort and image attachmentsPersonal
1.0Latest1 release · newest 20 August 2026

1.0 — everything

1.0.0-beta

Latest#

Added

  • NeuralVault Free. No account, no licence, no sign-up, and no expiry. Free is a chat-first build with three surfaces: Chat, the Vault browser, and Settings. Because Chat is its only destination, Free has no sidebar, and Chat runs full width. Cloud models are allowed on Free with your own API key, and local models through Ollama need nothing but Ollama.
  • Integrations are free and unlimited on every edition. Connect a curated MCP server and it lands what it finds into the vault as ordinary notes. Getting data in is free; the Dashboard that visualises it is Personal. - Not in this build: the six-source connector catalogue (Shopify, YouTube, Meta Ads, Search Console, GitHub, RSS). It exists in the previous Tauri app and has not been ported to this one. This line previously claimed those six sources by name, which was true of that app and is not true of this one.
  • Accounts and licensing. Sign in from Settings, and the app renews its own licence. Licences are device-bound, expiry is enforced at every gate, and you can see and end your running sessions.
  • Purchasing. Stripe checkout and subscriptions, per-seat pricing, organisations created at purchase, member invites, and team seat administration in the desktop app.
  • Session handoff. /nv-handoff writes a self-contained note that lets a conversation be resumed cold. Personal and above.
  • Connect OpenRouter by logging in rather than by pasting an API key.

Changed

  • Model routing now routes by the kind of work, not by model size.
  • Entitlements are enforced at the command layer, so a hand-crafted call cannot reach a gated feature. A test walks every capability and fails if any of them gates nothing at all.
  • Locked pages are hidden and route-guarded, so a deep link, a restored last page or a command palette entry cannot strand you on a page you cannot open.
  • Turning a paid feature off is deliberately never gated, so a lapsed licence cannot trap an install with sync running and the off switch refusing.

Fixed

  • Free no longer runs the paid agent fleet. The scheduler jobs checked no edition, so a Free install with a cloud key could dispatch unattended agent work and spend the user's own API credit without ever showing it. The jobs now refuse to start rather than starting and erroring.
  • Long local answers are no longer cut off. The Ollama client had a 120 second total request deadline on a streaming response, which severed ordinary long answers from a local model mid-sentence. It is now a 120 second idle deadline.
  • The plugin API is gated. The plugin host previously bound its socket and issued capability tokens on every edition. On Free it is not started at all.
  • A new install opens on the welcome screen instead of an empty "Select a session".
  • Onboarding stops selling what the edition cannot run, and adopts the Ollama models you already have instead of detecting and discarding them.
  • Embedding failures are surfaced. A missing nomic-embed-text was swallowed, so semantic recall silently did not work and notes were re-embedded forever against a 404.
  • Settings hides paid categories rather than rendering live controls that refuse on save.
  • The starter notes and the getting started guide describe the edition you actually have, including the fact that there is no Map page in any edition yet.

Security

  • A licence cannot be hand-edited in a shipped build, and an unsigned licence is refused.
  • A row-level security policy that would have exposed every organisation's subscription was fixed before release.
  • "Audit trail export" was removed from the tier table rather than left sold, because it does not exist. Reading your own audit log is free on every edition.
0.91 release · newest 1 August 2026

0.9 — everything

0.9.0-beta

Latest#

Added

  • Markdown vault as memory. Point the app at a folder and it reads and writes it in place, wikilinks and frontmatter included. No database, no import step, no lock-in.
  • Chat with Cortex, hybrid lexical and semantic search over the vault.
  • Agent fleet. Agents are Markdown files, not opaque config. Cron scheduler, autonomous board runner, and file triggers.
  • Board and Workflows. A kanban board agents can work, plus a workflow builder and skills.
  • Council, Squad and Gauntlet multi-model orchestration.
  • Approvals queue. Nothing is marked done, and no message is sent, without you.
  • Checkpoints and rewind, byte-exact, on every file write, plus an audit log and run ledger covering every tool call.
  • Local models via Ollama by default. Cloud providers are opt-in, and their API keys are held in the Windows Credential Manager, never in a plaintext file.
  • Dashboard, Vault browser and Activity views.
  • Watch and Learn screen recorder and Voice Memos.
  • Calendar and a self-rewrite review flow.
  • Plugin API, a line-delimited JSON protocol over a local socket, implementable from any language.
  • Sync to a sync server you run, with per-user accounts and folder-scoped permissions.
  • Business connectors: Shopify, YouTube, Meta Ads, Google Search Console, GitHub and RSS.
  • Wire, team channels and email, read side.
  • Legal documents shipped with the build: EULA.md, PRIVACY.md, THIRD-PARTY-NOTICES.md and LICENSE.
  • Lifecycle hooks, as of 2026-08-16. Run a command of your own at the start and the end of an agent turn. Configuration is a file, ~/.neuralvault-hooks.json, and it is off by default: an install that has never been configured runs nothing. Each hook names an event, turn_start or turn_end, and a command line, and can filter by run kind and by nesting depth. By default a hook fires for top-level turns only, not once more for every sub-agent a turn goes on to dispatch. Context reaches the command as environment variables (NV_HOOK_EVENT, NV_RUN_ID, NV_RUN_KIND, NV_RUN_DETAIL, NV_RUN_DEPTH, and on turn_end also NV_RUN_STATE, NV_TOKENS_IN and NV_TOKENS_OUT) rather than being substituted into the command text, so nothing you type into a chat message can become part of a command. Hooks run detached and are killed at a hard timeout, 30 seconds unless you raise it, so a hook that hangs cannot hold up a turn. Version 1 is observe-only: a hook is told what happened and cannot block, delay or veto anything. There is no pre-tool event for a hook to refuse an action at, and no settings screen yet, so the file is the interface. Personal edition. The design and its reasoning are in docs/superpowers/specs/2026-08-16-lifecycle-hooks-design.md.
  • Report a bug, as of 2026-08-16. Settings, About has a Report a bug button. It opens a prefilled issue form on the public tracker, github.com/SamuelNDCE/neuralvault-releases, in your normal browser. The application sends nothing itself: it opens a page, and nothing leaves your machine until you press submit on GitHub, where you can still edit or abandon the report. Before the browser opens, the app shows you the exact diagnostics text it proposes to attach (app version, operating system, WebView version, whether a vault is configured, which background jobs are on, and the tail of the crash log if there is one) and you can decline it and file the report anyway. That text never carries your vault path, note content, prompts or model outputs. The tracker is a separate repository from the source, which is private, and it is deliberately public, so treat anything you attach as readable by anyone. PRIVACY.md section 3.5 is the full statement.
  • Modular entitlements, as of 2026-08-16. Groundwork only: nothing is sold this way yet. A licence can now carry a list of individually granted capabilities on top of its edition, so a tier can be a floor plus the add-ons a customer actually wanted rather than one fixed bundle. The intent is that somebody working alone is not made to pay for team features they will never open. This changes nothing for anyone today: no tier uses it, the list is empty in every licence, and an empty list is byte-for-byte identical on the wire to a licence issued before the field existed, so nothing already issued is invalidated. Add-ons are strictly additive and can only ever grant, never remove, and only a validly signed licence can name them: an unsigned file cannot, even in a development build where unsigned files are otherwise honoured. A name the installed version does not recognise is ignored rather than treated as an error, so a licence issued by a newer server never locks a paying customer out of the product.
  • Commercial tiers are no longer tied to edition names, as of 2026-08-16. A tier now records which of the three built-in editions it starts from, so new tiers can be added without shipping a new build. Previously a tier whose name the installed application had never seen would have been rejected outright, and because licences are checked offline there would have been no way to explain why.

Fixed

  • EULA.md no longer carries an internal draft warning and an internal notes block addressed to the author. Both are displayed verbatim by the onboarding consent step and the installer licence page, so they moved to docs/legal/eula-review-notes.md, which never ships.
  • EULA.md section 7.3 referred to a privacy notice that did not exist. PRIVACY.md now exists and is distributed with the application.

Known issues

  • Windows only. The build targets the NSIS installer and nothing else. There is no macOS or Linux build.
  • The installer is not code-signed. Windows SmartScreen will warn on first run. This is expected for the beta and is being addressed. The signing route is decided (Azure Trusted Signing) and recorded in docs/release/code-signing.md; it is waiting on the certificate profile itself, which has days of lead time.
  • Auto-update can check, but it cannot install. Corrected 2026-08-16. This entry used to say the updater had nowhere to check and that there was no in-app update check at all. Both are now out of date. A release host exists: plugins.updater.endpoints points at github.com/SamuelNDCE/neuralvault-releases, a public repository that holds installers and update manifests and nothing else. It has to be a separate public repository, because the source repository is private and an update endpoint must be fetchable with no credential. Settings, About can run the check and tell you whether a newer version exists. What is still missing is the second half: nothing downloads or installs an update. download_and_install is deliberately not called from anywhere, and a test fails if it ever is, because the installers are not code-signed, so one fetched automatically could not be verified as having come from us. A new version is therefore still installed by hand. The signing half is tracked in docs/release/code-signing.md.
  • Wire cannot send. Wire's registered surface is read, draft, approve, dismiss and channel management. There is no outbound transport of any kind, so a drafted message cannot actually be sent from the app. Mailbox password rotation is likewise not reachable from the UI.
  • The Truth Engine is cut. Its Settings card is gone, and its contradiction scanner is not part of this release. If you were relying on the ignored-folders setting, it no longer does anything.
  • Editions are enforced, as of 2026-08-10. This entry said the opposite until 2026-08-12, and it was true when this heading was dated: the gate landed inside 0.9.0-beta without a version bump, so the notes and the build disagreed for two days. This is still a single codebase and a single build. Every install runs as Free until a licence file is present, and the check is at the command layer rather than in the interface, so hiding or showing a button does not change what an install may do. Enforcement is not yet complete: the Plugin API is ungated, the read side of several paid rows still answers on Free, and the unattended background jobs run regardless of edition. README.md carries the current list. One build ships every edition, so this is a commercial boundary against the caller inside the app, not protection against the owner of the machine editing a file in their own home directory.
  • Free is a chat-first build, as of 2026-08-12. Chat, Settings and the Vault browser, and only Chat is a destination, so a Free install has no sidebar: it is one chat page. The Board, the dashboard, the agent fleet, the Workflows builder, Skills, Watch & Learn, Voice Memos and Canvas need Personal. Local models through Ollama, checkpoints and rewind, the audit log and the Council, Squad and Gauntlet orchestration modes stay free. So do cloud providers: gating them was tried on 2026-08-12 and reversed the same day, because almost nobody has a local Ollama install ready on first run and a chat-first free build that cannot reach a model is not a product. You bring your own API key either way. The Editions table in README.md is the list, and edition_table_matches_readme now fails if the code and that table disagree.
  • Beta stability. Features may change or be withdrawn. Keep independent backups of your vault, as EULA.md section 9 says. Do not make this the sole store of anything you cannot afford to lose.
  • Uninstalling does not remove saved API keys. The uninstaller removes NeuralVault's settings files and leaves your vault alone, but it cannot reach the Windows Credential Manager, where API keys and connector credentials are held. To remove those, open Settings, General, Remove my data before uninstalling.
  • Crash reports are local only. A crash writes ~/.neuralvault-crash.log and stops there. Nothing uploads it, and nothing ever will on its own. It reaches us only if you send it to us yourself, or attach it to a bug report using the button described above, having read what it contains first.
  • A malformed hooks file disables every hook, silently, as of 2026-08-16. ~/.neuralvault-hooks.json is parsed whole: one bad entry, a trailing comma or an event name that does not exist, throws away the entire file and every hook in it, with no message. That is the safe direction for a file that names commands to run, and it is what every other ~/.neuralvault-* config does, but "my hooks stopped firing" and "I mistyped one line" look identical from the outside. Check the file is valid JSON before concluding the feature is broken.

Most entries apply to every edition. Where a release affected only one, that entry appears under that edition alone.

NeuralVault follows Keep a Changelog. Want these by email when they happen? Pick NeuralVault releases on the services page.