mimir-os 0.1.0

Mimir — the personal agent operating system. Self-hosted. Multi-agent. Auditable for every decision.

Mimir

The personal agent operating system. Self-hosted. Multi-agent. Auditable for every decision your team makes.

Mimir is the supervisor that lives on your VPS. He runs your agent turns behind a verifiable, Ed25519-signed audit chain — every decision inspectable, every turn replayable in principle. The design goal: agents whose every action passes an auditor's test.

You supervise Mimir. Mimir supervises the work.

Status: v0.1.0 — first public alpha (targets July 2026). Live now: the self-hosted auditable supervisor — mimir run "<task>" (a LOCAL, audited agent turn), mimir verify / mimir audit over the signed chain, tier-filtered vault reads, the MCP server, and a one-command Docker install.

Gated — on the roadmap, not in v0.1.0: multi-agent delegation (spawn / watch), skill training (train), audit branch / replay / export, and the multi-platform adapters. These ship after the delegation milestone; the CLI stubs print a "deferred" notice today. Code samples for them below are marked # GATED. Substrate (persistence-os) v0.8.5a1, 2093 tests passing.

---

Install

Paste this into Claude Code (once v0.1.0 ships):

/mimir install

Claude Code will provision Mimir on your VPS or local machine, mint your Ed25519 keys, wire up your MCP, and start the agent loop. 90 seconds, end to end.

→ Watch the install demo (1:32) — coming Week 12

Six things Mimir does

Items 1–4 are gated (roadmap, after the delegation milestone). Items 5–6 are live in v0.1.0. Gated samples are marked # GATED.

1. Spawns specialists under your authority. (gated — roadmap)

# GATED — not available in v0.1.0 (pending C3 delegation flip)
mimir spawn --scope "watch my GitHub PRs and summarize once per hour" --tier SAFE

2. Trains them from their audit logs. (gated — roadmap) When a child accomplishes something well, Mimir will distil the pattern into a skill, run it through four promotion gates (replay byte-identity, audit-chain integrity, score delta on held-out tasks, stub coverage), and register it in that child's library.

# GATED — not available in v0.1.0 (pending C3 delegation flip)
mimir train watcher-7af1

3. Coordinates them via software-transactional memory. (gated — roadmap) Two children that need shared state (scribe writing email-summaries, courier updating calendar from them) commit atomically — either both ship or neither does.

4. Forks-and-votes when uncertain. (gated — roadmap) Mimir will spawn N children to try N approaches, score them, and commit only the winner's facts — the N-1 losing branches leaving zero trace in your vault history.

5. Filters every read by tier. (live) Each turn carries a Mom-tier capability token (PUBLIC / SAFE / SENSITIVE / PII). Vault queries are filtered server-side; a SAFE-tier read literally cannot reach PII datoms even if it tries.

6. Audits everything end-to-end. (live) Every turn, every read, every write, every halt — one Ed25519-signed row in a hash-chained log. Verify it yourself:

mimir verify
mimir audit

Audit replay / branch / export are gated (they ship with the delegation milestone):

# GATED — not available in v0.1.0 (pending C3 delegation flip)
mimir replay --agent watcher-7af1 --from chain://event-1247
mimir branch --agent watcher-7af1 --from chain://event-1247
mimir export --format=eu-ai-act --from 2026-05-01 --to 2026-08-01

A first-in-class memory graph

Bitemporal datoms (Datomic-style) underneath; an Obsidian-grade graph view on top. Search by tag, full-text, vector proximity, or graph distance. Tier × bucket taxonomy with biomimetic decay so old context doesn't drown new context. Mom-tier security classification (PUBLIC / SAFE / SENSITIVE / PII) on every memory.

Multi-platform reach (gated — roadmap)

The plan: talk to Mimir from WhatsApp, Slack, Telegram, iMessage, or email — same memory, same audit chain, every platform. The adapter surface is not in v0.1.0:

# GATED — not available in v0.1.0 (adapter surface lands in a later release)
mimir adapter add whatsapp --twilio-sid AC...
mimir adapter add slack --workspace acme.slack.com

Pricing

Tier	Price	What you get
Self-hosted	Free / Apache 2	The live v0.1.0 supervisor. Run the Docker image on your VPS or laptop.
Mimir Pro †	$29/mo	Hosted Ed25519 retention (90d), email support, multi-device sync, web UI access.
Mimir Team †	$199/mo (5 seats)	Shared memory graph, RBAC, team audit dashboard, 1y retention.
Mimir Compliance †	$2,500/mo	EU residency, ISO/IEC 24970 mapping, expert-witness export, dedicated tenant, SLAs, SSO.

† Planned — not yet available in v0.1.0. The hosted/managed tiers are on the roadmap and are not offered today; only the free Self-hosted tier is live. The OSS tier is the funnel; the Compliance tier is the intended path to fund development.

Three things only Mimir is designed to do (gated — roadmap)

The substrate makes these possible; the CLI surface ships after the delegation milestone (today the verbs print a "deferred" notice):

# GATED — not available in v0.1.0 (pending C3 delegation flip)
# 1. Branch any audit point at datom granularity
mimir branch --agent watcher-7af1 --from chain://event-1247

# 2. Replay any decision deterministically
mimir replay --agent watcher-7af1 --from chain://event-1247 --intervene tool=write_file:/etc/foo

# 3. Export an EU AI Act Article 12 evidence package
mimir export --format=eu-ai-act --from 2026-05-01 --to 2026-08-01 > evidence.tar.gz

Hermes-agent (Nous Research, 129K stars) only supports /undo and cannot do any of these without a substrate rewrite. Letta, Mem0, and AnythingLLM have memory but no audit chain. Cursor, Cline, and Aider have neither memory nor audit at the substrate layer.

Why Mimir exists

Built by Nawfal Saadi — Director of Business Development MENA & Green Hydrogen at Scatec ASA, running a multi-billion-dollar megaproject. After 9 months of 9pm-to-4am evenings in Casablanca building the substrate that would let his own team's agents survive cross-shift handoffs without losing context, Mimir is what came out.

The duality of operator + builder is the thesis: agents that pass an auditor's test deserve to be in production. Agents that don't, don't.

In Norse myth, after Mimir's death, Odin preserves his head and consults it as an oracle for every major decision. That's the relationship you have with your agents now.

Architecture

Mimir bundles four open-source projects under one Docker image:

• persistence-os — the bitemporal substrate (datom log, plan AST, audit chain, DB.fork, replay engine). AGPL-3.
• ai-box memory graph — tier × bucket vault with biomimetic decay and Mom security tiers. (Source published Phase E.)
• juba-bridge multi-platform — WhatsApp / Slack / Telegram / iMessage / email adapters. (Source published Phase E.)
• persistence-coder — the agent loop that powers Mimir and his children.

Mimir itself is the supervisor + capability-tier filter + skill curator + CLI surface that ties them together.

→ Architecture deep-dive — coming Week 12

Roadmap

Phase	Target	Deliverable
A	Week 1	persistence-os pushed to public origin; mimir-os repos seeded; namespaces locked.
B	Weeks 2-3	Ed25519 audit signing on substrate; tamper-evident demo video.
C	Weeks 4-6	persistence-coder MVP v0.9.0a1; wrong-path-steering demo video.
D	Weeks 7-11	Mimir bundle: supervisor, tiered vault, install skill, agent fleet UI, six CLI verbs.
E	Weeks 12-13	Public launch: HN, ProductHunt, PyPI release, Docker image, install command.
F	Week 14	Mimir Compliance tier: EU AI Act Article 12 export; first paid pilots.

License

Mimir (this repo) is Apache 2.0. The substrate underneath (persistence-os) is AGPL-3 — copyleft, runtime use unencumbered, only forks-into-closed-products restricted. Commercial license available for vertical integrators.

pip install mimir-os                    # PyPI
docker pull mimiros/mimir-os:latest     # Docker Hub (registry namespace differs)
/mimir install                          # Anthropic Skill, in any Claude Code session

---

Substrate: nawsaafa/persistence-os · Docs: mimir-os.com · Contact: [e-mail]