# Where ALINA is going: an AI that is yours

> ALINA is becoming an always-on intelligence personal to one human. It knows you, works across your machines in your name, and calls the strongest frontier models when a task needs them. Your memory, your authority and your identity live in an operating system you own. Models and agent apps are parts you can swap.

Part of [ALINA OS](https://alinaos.com/). Direction set October 2026. Status today: Loom ledger live · shared memory live · Engram in research · ALINA box is the target.

## The inversion

Today, everywhere, a frontier model is the agent and your history is a tool it calls, stored in its vendor's app. ALINA turns that over: a personal model that already knows you holds the plan, and frontier models are tools it calls for heavy reasoning. The change happens one role at a time, and only when the new arrangement wins on evidence.

## Hands, brains and a self

- **Hands (the harnesses):** OpenClaw, Claude Code and Desktop, Codex, pi and Antigravity, joined by [Loom](https://alinaos.com/loom/) so any of them can see, follow and take over another's session. No vendor's app is the hub.
- **Brains (the models):** frontier models from several labs today. A personal model, Engram, joins as one more brain and takes over roles only as it earns them.
- **Self (owned by the OS):** memory, authority and identity. It outlives the hands and the brains.

## Three memories

| Kind | System | Holds | Status |
|---|---|---|---|
| Episodic | Loom | Every session in every agent app, with provenance and trust labels; more than 150,000 events across five harnesses | Live |
| Semantic | Shared memory | Durable facts, decisions and lessons distilled nightly by a council of models from three labs; every read and write bound to an audience | Live |
| Parametric | [Engram](https://github.com/alinaos-org/engram) | Habits, style and ways of working in a personal model's weights; encrypted, detachable, rebuildable | Research |

No new memory store gets built: the ledger and shared memory are the retrieval log beside Engram's weights.

### The sleep cycle

Sessions → Loom (episodes) → memory council → shared memory (facts) → Engram consolidation → personal adapter. Steps up to shared memory run nightly today; consolidation is in research.

## Earned, never assumed

A personal model takes over a role only when it beats frontier models plus retrieval on a benchmark built from the person's real sessions.

| Role | Promoted when |
|---|---|
| Memory oracle | It beats frontier plus retrieval at recalling, updating and forgetting what you know |
| Voice | You prefer its drafts in a blind comparison |
| Routine conductor | It escalates correctly on held-out tasks, with no worse outcomes at lower cost |
| Conductor | A sustained routine-conductor record, with a frontier fallback kept |

The design is hybrid permanently: weights for who you are and how you work, retrieval for exact facts.

## Your data, your rules

| Data class | May train your model |
|---|---|
| personal | Yes |
| client | Never |
| export-controlled | Never |
| research-open | No (its own corpus) |

- Inside your systems first: ALINA changes your own machines, repositories and agents through reviewed paths.
- Outbound waits for you: email, chat, posts and purchases are drafted and sent only after approval, from the desktop or phone.
- Untrusted text stays data: web pages, tool output and other people's messages are labelled at the source.
- Every action has an author in the ledger.
- Trust widens one class of action at a time, on a record of approvals that needed no edits.

## The ALINA box (target)

Each person gets their own ALINA box: a NixOS machine at home running their agent apps, ledger, memory and personal model. The cloud is used for training bursts. A second person's ALINA is a build from the same flake, not a hand-made copy. Every layer is designed per person from day one.

## ControlStackAI Silicon (long range)

[ControlStackAI Silicon](https://github.com/controlstackai-silicon) works on agent-directed chip design: small RISC-V extensions and low-power, fixed-latency accelerators, RTL to FPGA to silicon. It meets ALINA in three places, in order: ALINA runs the design-verify-measure loop as its flagship workload; a wake-gating front end lets an always-on AI wait at almost no power; later, a unit for a personal model's fast memory updates.

## The road

| # | Milestone | Gate | Status |
|---|---|---|---|
| M0 | Loom ledger and presence, shared memory, frontier ALINA | Live in daily use | Done, Oct 2026 |
| M1 | Data classes in the ledger and memory | Client and regulated threads provably excluded from personal | Next |
| M2 | Loom takeover across all five harnesses | A session moves between any two apps with its handoff, under a lease | Next |
| M3 | Personal benchmark built from the ledger | Frontier plus retrieval baseline measured | Next |
| M4 | Engram as a provider on a canary agent | A week without regressions | Planned |
| M5 | Engram as memory oracle | Beats the M3 baseline on recall, update and forget | Planned |
| M6 | Authority policy enforced, outbound approval flow | Every outbound action approved and audited | Planned |
| M7 | Chip research loop as an ALINA workload | One full cycle reproduced from a flake | Planned |
| M8 | Engram as voice, then routine conductor | Role gates above | Planned |
| M9 | First ALINA box runs a personal model locally | Always on; cloud only for training | Planned |
| M10 | A second person's ALINA from the same flake | No hand edits; isolation verified | Planned |
| M11 | Wake-gating front end on FPGA | Power and latency measured against a CPU or GPU baseline | Planned |

## Principles

1. Keep today's capability until a replacement wins, one role at a time.
2. The OS owns the self.
3. Earn every role on the person's own work.
4. Per person from day one.
5. Evidence over claims.
6. Declarative and reproducible: a second ALINA is a build.