memforge-ai 0.2.0b3

The third beta of MemForge, with the memory plugin 0.2.0-beta.3 for Claude Code and Codex.

Jira sources default to a personal access token. A new Jira source authenticates with a personal access token, so it needs no local daemon. Browser session stays available, and local sync always uses it. Migration 133 writes each stored Jira source's existing method into its configuration, so every existing source authenticates as before.

Agent-session storage is removed.

• On upgrade, MemForge retires every remaining Codex Session and Claude Code Session Source as Source deletion does, and drops the agent-session tables.
• If such a Source still has open work, such as a queued or failed access change, the upgrade stops and names it. Restore the recovery point with memforge backup restore <id>, resolve that work on 0.2.0b2, delete the Source in the admin UI, then upgrade.
• POST /api/v1/agent-sessions/windows and POST /api/v1/hooks/receipts now answer 404.

Upgrade the plugin. If your coding agent shows MemForge hook skipped: /hooks/receipts returned HTTP 404 (or 410), update the plugin and restart the client:

• Claude Code: claude plugin marketplace update memforge, then claude plugin update memory@memforge.
• Codex: codex plugin marketplace upgrade memforge, then codex plugin add memory@memforge.

Document updates no longer wait for a person. No document update opens a Review any more:

• Two incompatible claims that both still have source text stay active, and relation discovery, which now also compares memories of one document, relates them as updates or contradicts.
• A revision replaces an old memory only after a separate check confirms the new claim states every fact of the old one; otherwise the new claim is added beside it.
• Sources no longer have a lifecycle gate, and an expired Review can no longer be rechecked. Reviews opened by earlier versions stay in the queue and can still be approved or rejected.

Breaking API changes.

• Removed: password sign-in (POST /api/v1/auth/login). memforge init no longer asks for an admin password, and migration 132 drops the users table.
• Removed: POST /api/v1/sources/{source_id}/memory-lifecycle/reviews/{review_id}/approve and /reject; use POST /api/v1/memory-reviews/{review_id}/approve or /reject.
• Removed: GET /api/v1/recent-changes, GET /api/v1/stats/detailed, POST /api/v1/memory-reviews/{review_id}/refresh and POST /api/v1/sources/{id}/memory-lifecycle/gate.
• source_filter.clients is removed from the search API.

Also in this release

• The admin UI is served at /. The previous admin UI moves to /v1/ for this release only, and links under /v2/ redirect to the same page at /.
• Relation discovery orders two statements recorded on the same day by their exact source time.
• Incremental extraction and Support revalidation decide what changed by what a model reads, so storage-only edits and moved Markdown sections no longer send unchanged content to the model.
• MEMFORGE_LLM_METADATA_MODELS lets a route LiteLLM has no limits for, such as a gateway alias, take the token limits of the model it serves.

Upgrading from 0.2.0b2. Before a release applies schema migrations, it saves a recovery point of the database and the vector index under backups/ in the data directory. memforge backup list and memforge backup restore ID list and restore recovery points. See the quickstart's Upgrading section for the Docker commands.

Install the package with pip install --pre memforge-ai, or follow the quickstart for the Docker stack. See CHANGELOG.md for the full list.

memforge-ai 0.2.0b2

The second beta of MemForge, with the memory plugin 0.2.0-beta.2 for Claude Code and Codex.

Automatic agent-session capture is removed.

• The plugins no longer upload session windows. Their only hook is SessionStart, which adds MemForge usage guidance and makes no network call.
• Agents write memory only when you ask them to remember, correct or retire something, through the confirmed create_memory tool.
• POST /api/v1/agent-sessions/windows and POST /api/v1/hooks/receipts answer 410 agent_session_capture_removed in this release, and the next release removes them.

Upgrade the plugin. If your coding agent shows MemForge hook skipped: /hooks/receipts returned HTTP 410 at the end of a turn, update the plugin and restart the client:

• Claude Code: claude plugin marketplace update memforge, then claude plugin update memory@memforge.
• Codex: codex plugin marketplace upgrade memforge, then codex plugin add memory@memforge.

You can delete the old queue files in ~/.memforge-agent (queue.sqlite*, worker.lock). Keep artifacts/ if you use get_resource(mode="file"). To drop existing agent-session memories now, delete the Codex Session and Claude Code Session Sources in the admin UI. Otherwise the next release retires them on upgrade.

Also in this release

• The admin UI V2 sets up and edits Sources on its own pages, drawn from the server's form description. A new config-check route previews derived values, and refused configurations answer config_invalid with field issues.
• A Notion connector.
• Teams conversation sync is more correct: page certificates, bounded backfills, and scope changes that supersede open transitions.
• The quick quality evaluation covers document updates and reconciliation.

Upgrading from 0.2.0b1. Before a release applies schema migrations, it saves a recovery point of the database and the vector index under backups/ in the data directory. memforge backup list and memforge backup restore ID list and restore recovery points. See the quickstart's Upgrading section for the Docker commands.

Install the package with pip install --pre memforge-ai, or follow the quickstart for the Docker stack. See CHANGELOG.md for the full list.

memforge-ai 0.2.0b1

The first beta of MemForge. It runs for one user on one machine from the Docker Compose stack, and you choose the models in Settings.

Not supported yet: upgrading an installation, and backing up or restoring its data. Start the beta on a fresh memforge-data volume. If an upgrade fails, start again on a fresh volume and sync your sources again.

Highlights since 0.1.58

• Model settings: choose a provider and model in Settings, list models by kind, and take token limits from the provider's list.
• Closed-choice model calls answer without reasoning first, so a model that reasons by default no longer uses up their output budget.
• Model listing ends at its timeout.
• Search includes your own private memories by default, including from the CLI and the API.
• Quickstart guidance for setting up models and for reaching providers through a proxy.

Install the package with pip install --pre memforge-ai, or follow the quickstart for the Docker stack. See CHANGELOG.md for the full list.