What shipped, when.
Milestones only, condensed from the product changelog. Every release is on GitHub.
- 0.2.3September 2, 2026
The agent docks beside the page
The Mari agent sits beside the page on desktop instead of floating over it, so the content narrows to make room, and the conversation survives navigation between pages. Published knowledge chats stay focused and never draw a second agent. The verification audit on Facts renders through the shared table so it matches the claims ledger. Helm values can set node selectors, affinity, tolerations, and topology spread constraints per workload. Helm chart 0.2.3 with no migration.
- 0.2.2September 2, 2026
Knowledge chat searches everything, and private Slack by channel
Knowledge chat searches every document its destination can see rather than only approved answers, which stay a reviewed fast path for repeated questions. Answer context grows from four documents to eight, a learned workflow can guide an answer without replacing the question that was asked, and a grounded answer keeps its evidence even when the model opens hesitantly. Slack documents carry their channel id and name, a question about a private channel resolves to that exact channel newest first, and Mari's own replies and Slack system events stay out of the knowledge base. Helm chart 0.2.2 with one additive migration.
- 0.2.1September 2, 2026
Incremental fact scans and a published Helm chart
Fact extraction reads each passage once per content hash, so a rescan mines only what changed, and claims that differ only in case, spacing, or punctuation collapse before they reach review. A dedupe script merges historical duplicates and keeps their provenance. Pushing a release tag now builds the images, publishes the Helm chart to GitHub Container Registry, and attaches the chart and its checksum to the release. The Slack app manifest subscribes to private channel messages, and a knowledge chat destination is readable the moment it is created instead of opening as a half-drawn card. Helm chart 0.2.1 with one additive migration.
- 0.2.0September 2, 2026
Fact intelligence, scheduled tasks, and a hardening pass
Facts carry embeddings and cluster in time, extraction runs on a budget with staged review, and invalidated facts can be restored. Scheduled tasks get their own page with a heartbeat that fails a dead run instead of blocking the schedule. Sources can be edited, removed, or connected twice, and every source reconciles in full once a day so deleted pages leave search. Admin rights come from project membership, public chat sessions are token bound and rate limited, and Slack tells you at connect time when the app cannot see a channel. Helm chart 0.2.0 with nine additive migrations.
- LaunchAugust 11, 2026
Open source, Apache 2.0
Mari is open source and self-hosted by design: Postgres with pgvector is the whole data plane, local models run through ollama, and every LLM feature has a deterministic fallback. Ships with 12 connectors, cited answers, facts with owners and freshness, the lineage graph, and the drift audit.
- ConsoleAugust 10, 2026
Founder accounts and the live demo
The hosted console is limited to seeded founder accounts. The public demo at mari.guru/demo runs the same console on demo data, and the demo Lambda can pin a generation model at boot.
- 0.0.1July 26, 2026
Desktop app
A self-contained desktop app for Apple Silicon, so a single person can run the whole product without Docker.
- LicenseJuly 25, 2026
Apache 2.0
The repository adopts the Apache License 2.0 ahead of the public launch, later paired with a contributor license agreement for outside contributions.