# Lotti
[](https://codecov.io/gh/matthiasn/lotti) [](https://flathub.org/en/apps/com.matthiasn.lotti) [](https://github.com/matthiasn/lotti/releases) [](LICENSE)
Come say hi on [Discord](https://discord.gg/uuSaa8NpY) — questions, support,
ideas, or just what you are using Lotti for.
**A distributed system of record for your memories — kept on your devices,
and nowhere else.**
Everything you capture in Lotti is a memory: a voice note, a screenshot, a
conversation, a photo from a birthday party, an hour of focused work, the
journal entry you spoke into your phone last night. Lotti keeps those memories
on your own devices and syncs them between those devices with end-to-end
encryption; no cloud ever holds a readable copy. On top of that record sit
applications — tasks, relationships, goals, events, projects, a daily journal —
most of them looked after by a personal AI agent that reads what you recorded and
proposes what to do next. Agents propose; you decide. Your memories are not
rewritten behind your back, not even by the agents that work for you.
macOS · Linux · Windows · iOS · Android. Flutter and Dart, GPL-3.0, in
development since 2016.
I have tracked around 11,000 hours of my own work in it since 2022.
<!-- All screenshots come from the manual's deterministic fixture pipeline
(docs-site/metadata/screenshot-cases.json), so they track the build. They
point at the `development` channel and update when the manual regenerates;
if a case id is ever renamed, these links need renaming with it. -->
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://pub-3df7bcf4b8ca493fa6acea182d69d9c7.r2.dev/manual/screenshots/development/tasks/workspace/desktop-dark.webp">
<img alt="The task workspace: a filtered task list beside an open task with its cover art, status, labels, and AI summary" src="https://pub-3df7bcf4b8ca493fa6acea182d69d9c7.r2.dev/manual/screenshots/development/tasks/workspace/desktop-light.webp">
</picture>
[Read the manual](https://matthiasn.github.io/lotti/manual/development/) ·
[Install](#install) ·
[Blog series](https://matthiasnehlsen.substack.com/p/meet-lotti)
---
## The vision: a system of record you own
Most personal software treats your life as content to be stored on someone
else's computer and mined. Lotti starts from the opposite premise: your
memories are a *record*, and a record is only worth something if it is
complete, unaltered, and yours. Four principles follow from that.
### 1. Your memories live only on your devices
There is no Lotti cloud, no Lotti account, and no server that stores your
memories. Each device holds the full record in a local SQLite database, with
audio, images and other attachments beside it on the filesystem. No telemetry,
no analytics, and nothing uploaded to Lotti — you can confirm that by reading
the source or watching the traffic.
The record is *distributed*, not centralised: every device you pair is a full
peer. Pairing gives a new device everything written from then on; once you send
it your history from a device that has it, a phone, a laptop and a desktop each
carry the whole record, and losing one costs you nothing as long as another
survives. Your data is a file you
already have, with a documented schema, so there is no export wizard and
nothing to unlock.
### 2. End-to-end encrypted sync through a zero-trust relay
Devices talk to each other through a relay rather than a storage service: a
[Matrix](https://matrix.org) homeserver (Synapse). The relay is a zero-trust
component. Whoever operates it — you, a friend, a hosting company — never holds
a key, so they cannot read your memories, and they cannot slip changes into
your record either. You do not have to trust the operator with your data; you
only rely on them to keep the relay running.
```mermaid
flowchart LR
subgraph D1["Your laptop"]
L1[(Local SQLite<br/>+ attachments)]
end
subgraph D2["Your phone"]
L2[(Local SQLite<br/>+ attachments)]
end
R{{"Matrix homeserver<br/>zero-trust relay<br/>(ciphertext only)"}}
L1 -- "Megolm-encrypted events<br/>AES-256-CTR attachments" --> R
R -- "ciphertext delivered;<br/>keys only to verified devices" --> L2
L2 -- "the same, in reverse" --> R
R --> L1
```
How it works:
- **Encryption.** Each change is sent as an event in a private, non-federated
sync room created with Matrix's `m.megolm.v1.aes-sha2` encryption. Room
messages are encrypted with Megolm, and the Megolm session keys are exchanged
between devices over Olm, using the [Vodozemac](https://github.com/matrix-org/vodozemac)
implementation of both. Audio and images are uploaded as Matrix encrypted
attachments: AES-256-CTR with a per-file key, IV and SHA-256 hash that travel
only inside the encrypted event.
- **Only verified devices get keys.** Keys are shared only with devices you
have verified yourself through an emoji comparison. A device that joins the
account without completing verification receives ciphertext it cannot read,
and incoming events from unverified devices — or anything sent in plaintext —
are dropped rather than applied, so whoever operates the server cannot inject
changes into your record.
- **Trust comes from verification, not from the server.** The emoji
comparison is what stops a relay from posing as one of your devices: verify
only devices you are holding, and the operator never gets a key.
- **What zero trust does not cover.** The operator never sees plaintext, but
can see metadata: the account ID and room name, which device synced and when,
IP addresses, device display names, and the size of events and files. They
also control availability — they can delay, withhold or delete ciphertext,
but not read or alter it, and your devices still hold the full record.
- **The relay is not an archive.** Sync rooms carry a 30-day
`m.room.retention` policy. Synapse only enforces that when the operator turns
retention on (and media retention separately); the hosted Lotti Sync
service runs its own purge as well. Either way, the durable copy of your record is on
your devices, not on the relay.
- **Conflicts are resolved, not overwritten.** Every entry carries a vector
clock keyed by device, so a device that was offline for a week rejoins
without losing writes, and genuinely concurrent edits surface as conflicts
for you to resolve. Devices also track each other's counters and ask a peer
to re-send anything missing.
The full runtime picture — pairing, the sync log, backfill and failure handling —
is in the [sync concept](knowledge/features/sync/overview.md) and
[security and privacy](knowledge/architecture/security-and-privacy.md).
### 3. Memories stay unaltered, with clear provenance
A system of record is worthless if it can be quietly edited. Lotti keeps *what
you said and did* and *what an agent thinks* in separate databases on disk, and
agent-authored changes reach your record through a code path that requires
your approval: an agent suggests, and nothing happens until you confirm. A
confirmed checklist change keeps an approval receipt, and the writes an agent
makes without asking are logged in its audit trail.
A small, named class of writes is pre-approved rather than proposed, because
asking every time would add friction without telling you anything new — for
example transcribing a recording, saving an AI summary as a *new* entry beside
your own, or titling a task that has none. The exact set is defined in
[ADR 0102](docs/adr/0102-pre-approved-agent-changes.md) and can only grow by
another decision record. Some things are human-only by construction: an event
agent has no tool that can set an event's rating or cover photo.
**Record provenance is work in progress.** The goal is that every entry carries
a verifiable record of who authored it — you, an agent acting under a consent
you gave, or an agent proposal you approved — signed and chained per device so
that the record can prove it was not altered. The cryptographic building
blocks are [in the codebase](knowledge/features/provenance.md) and modelled in
TLA+, but entries are not signed yet, and the design will keep moving as this
lands.
See [Two databases](#two-databases-human-in-the-loop-by-construction) for how
the separation is enforced.
### 4. Inference that keeps nothing — and eventually runs on your own devices
Agents need a model, and today the most capable models mostly run in a
datacenter. Lotti has no inference backend of its own: every request goes to a
provider you configured, under your own key, routed per category. Ideally that
is a provider with a **zero-data-retention** policy, one that processes the
request and keeps neither the prompt nor the response, so that inference leaves
no copy of your memories behind. Lotti cannot verify a provider's retention
terms; choosing one is your due diligence, covered under
[Your inference is a routing decision](#your-inference-is-a-routing-decision).
The long-term direction is inference on your own devices, as they become
powerful enough to run capable models. The record and the reasoning over it
would then both stay at home. Part of that is already here: speech recognition
runs fully offline with Whisper or Voxtral, and lighter jobs run on a local
model. For the agents, today's answer is the cloud with zero data retention,
because they need a larger class of model than personal hardware runs fast
enough yet — see
[Local models and the sweet spot for agents](#local-models-and-the-sweet-spot-for-agents).
---
## Applications on top of the record
The storage layer is the product's foundation; the applications are what you
use every day. Each one reads from the same record, and all of them except the
journal can be given a persistent AI agent with its own report, memory, wake
schedule and proposal history. Tasks and the journal are on by default; the other applications are
**Sections** you switch on under Settings → Sections. Turning a section off
only hides it — nothing you recorded is deleted.
### Tasks, with task agents and pull request tracking
Every screenshot and audio recording is a memory, attached to the work it
belongs to. A rambling voice note comes back as a task with a checklist.
A **task agent** reads everything linked to a task and keeps a living report.
It proposes changes rather than making them: a title, an estimate, a due date,
priority or status, new or ticked-off checklist items, labels, follow-up tasks,
links to related tasks, time entries. Each proposal waits for you to confirm or
dismiss it. Automatic updates decide when the agent wakes, not what it may
change.
<p align="center">
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://pub-3df7bcf4b8ca493fa6acea182d69d9c7.r2.dev/manual/screenshots/development/tasks/agent-suggestions/mobile-dark.webp">
<img alt="A task agent's report with two proposed changes, each with a dismiss and a confirm control, plus Confirm all and the automatic-updates toggle" src="https://pub-3df7bcf4b8ca493fa6acea182d69d9c7.r2.dev/manual/screenshots/development/tasks/agent-suggestions/mobile-light.webp" width="380">
</picture>
</p>
**Pull request tracking** closes the loop between what you planned and what
actually shipped. Turn it on for a task from its "+" menu, link one or more
GitHub pull requests, and Lotti keeps a summary of each — a one-liner and a
TL;DR — refreshed when you open the task or ask it to. The pull requests become
part of the task agent's context, so it can propose ticking off the checklist
items a merged PR actually covers, naming the PR as its evidence. You still
confirm each one. A personal access token, kept in the device keystore and
synced end-to-end encrypted, is only ever sent to `api.github.com`.
**Coding prompts** and other briefs are a skill you trigger: they turn a task
plus your notes into a coding, design, image or research prompt you can paste
into the tool of your choice.
### Relationships, with a relationship agent
The People section is for the relationships you want to look after. Record a
conversation, or dictate afterwards who you spoke to and what you talked about;
a check-in holds the recordings, transcripts, notes and photos.
Mark someone as important and they get a **relationship agent**. Before you
meet again it gives you a briefing: what was discussed last time, how things
have been going, suggested talking points, and what to pay attention to or
steer clear of. It reminds you when it has been too long, and it can propose up
to three tasks arising from a conversation — each waiting for your
confirmation.
### Goals, with goal agents
Define a goal — hours on a category or label, a measurement you track — and a
**goal agent** watches your actual record against it. A deterministic evaluation runs
every day at no inference cost. When you drift off track, the agent reminds
you, with an OS notification and a banner explaining what slipped; when you are
on track, it stays out of your way, its report records the progress, and you
can ask it for an encouraging nudge. If the goal itself no longer fits, it can
propose a revision, which you approve.
### Events
Track the events that matter — a trip, a birthday party, a gathering — with the
photos, recordings and notes that belong to them. An **event agent** narrates
what you captured into a short living recap. The star rating and cover photo
stay yours alone: the agent has no tool that can set them.
### Projects
A project groups the tasks that serve one outcome. A **project agent** keeps a
report on what is going on: current status, a health verdict, what is still
missing, and recommended next steps. It can propose a status change or a new
task for your approval. The project's summary also flows into every task
agent's context, so work on a single task is informed by the project it
belongs to.
### Daily journal
Speak or type a journal entry, attach photos, or let a day's recordings stand
as they are. Transcription runs locally with Whisper or Voxtral if you want it
to, and your journal can be routed to a local model — or to no model at all.
The journal is the purest form of the record: your words, kept as you said
them, on your devices — read by no model unless you route it to one.
### You choose the brain, and you can see what it cost
Lotti has no inference backend. Route each category of your life to the compute
you are willing to stand behind: a cloud provider with zero data retention,
the European option Lotti recommends, or a local model as personal hardware
catches up. The usage
view reports tokens and requests for every cloud call, and spend, energy and
CO₂e for the providers that report them — today that means Melious. Local
inference is not measured at all, because the cost moves onto your own hardware
and grid.
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://pub-3df7bcf4b8ca493fa6acea182d69d9c7.r2.dev/manual/screenshots/development/ai/usage/desktop-dark.webp">
<img alt="Usage & Impact: cost, energy, CO2e, tokens and requests for the month, with cost broken down per day and per category" src="https://pub-3df7bcf4b8ca493fa6acea182d69d9c7.r2.dev/manual/screenshots/development/ai/usage/desktop-light.webp">
</picture>
### The sync protocol and the agent runtime are formally verified
Keeping several devices consistent without a server in charge is a
distributed-systems problem, and it is treated as one. The parts where
concurrency and crashes could lose or duplicate your data — the sync log, how
agents are woken and recover after a crash, how a confirmed suggestion is
applied when two devices race, how synced agent state converges, the agents'
message log, pull request tracking, and the Daily OS jobs — are written down as
[50 TLA+ models](specs/tla/README.md) and model-checked with TLC every night
across 177 configurations, under crashes, injected failures, clock skew and
messages arriving in any order ([the ledger](specs/tla/LEDGER.md) keeps the
running totals). No saved change is ever declared lost, and as long as devices
keep reconnecting and a failed send is retried, every saved change reaches every
device. A double tap never applies a confirmed change twice. Checking the models
found and closed dozens of real bugs. These are statements about the design, not
a proof that every line of code matches it; [the specs](specs/tla/README.md) say
exactly what is covered and which cases remain open, and generated traces drive
the real code through the same scenarios. The code is held to them the ordinary
way, too: over 35,000 tests at 99.9% line coverage, including more than 900
property-based tests.
---
## Install
| Platform | Where to get it |
|--------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------|
| **Linux** | [Flathub](https://flathub.org/en/apps/com.matthiasn.lotti) (recommended), or AppImage or `tar.gz` on [Releases](https://github.com/matthiasn/lotti/releases) |
| **Android** | APK on [Releases](https://github.com/matthiasn/lotti/releases), or Play Store internal testing (limited; invitation only) |
| **macOS** | Signed and notarized DMG on [Releases](https://github.com/matthiasn/lotti/releases) |
| **iOS / iPadOS / macOS** | TestFlight (limited; invitation only), with broader availability planned |
| **Windows** | Build from [source](docs/DEVELOPMENT.md) for now |
[](https://flathub.org/en/apps/com.matthiasn.lotti)
**AppImage (Linux, no app store):** download `Lotti-<version>-x86_64.AppImage`
from [Releases](https://github.com/matthiasn/lotti/releases), mark the
downloaded file executable (`chmod +x` on it, or your file manager's
permissions dialog) and start it. It brings its own media, keychain and
recording libraries and runs on distributions from 2022 on
(glibc 2.35 or newer: Ubuntu 22.04, Debian 12, Fedora 36 and later) that have
GTK 3 and Mesa's OpenGL ES, as any GNOME or KDE desktop does. It mounts itself
with FUSE, which desktops ship by default; where FUSE is missing, run it with
`--appimage-extract-and-run`. Unlike Flathub, it does not update itself.
---
## Features at a glance
### Capture
- **Audio recording** anywhere in the app, transcribed locally with Whisper
(99 languages) or Voxtral, or through a cloud provider with audio support.
- **Entries**: notes, images, screenshots, measurements, and surveys, attached
to the work, person, event or day they belong to.
### Organize and reflect
- **Tasks** with full lifecycle (open, groomed, in progress, blocked, on hold,
done, rejected), checklists, estimates, priorities, due dates, labels, linked
context, optional generated cover art, and linked GitHub pull requests.
- **Time tracking** recorded against the plan rather than instead of it, with
focus ratings. A task describes the outcome you want; a time record describes
what actually happened. Lotti keeps them as separate facts, so the record
stays honest when the week gets noisy.
- **Time analysis** over categories, habits, and measurements.
<picture>
<source media="(prefers-color-scheme: dark)" srcset="https://pub-3df7bcf4b8ca493fa6acea182d69d9c7.r2.dev/manual/screenshots/development/time-analysis/overview/desktop-dark.webp">
<img alt="Time Analysis: total, focused and other hours for the month, time per day stacked by category, and a per-category table with share and daily average" src="https://pub-3df7bcf4b8ca493fa6acea182d69d9c7.r2.dev/manual/screenshots/development/time-analysis/overview/desktop-light.webp">
</picture>
- **Categories and labels** to make the boundaries that your decisions
actually use.
- **Habits, measurables, and health data** imported from Apple Health and
other sources.
- **Sections** you switch on under Settings → Sections: Daily OS, Projects,
Goals, Habits, Dashboards, People and Events. They are in daily use and ship
switched off, so you choose which parts of Lotti you use.
### AI and automation
- **Agents for every application**: task, project, goal, relationship and event
agents, plus a day planner. Each keeps its own report, memory and proposal
history.
- **Providers, models, and inference profiles** as three separate layers, so
routing is explicit and the blast radius of a change is visible before you
make it.
- **A European route, offered rather than imposed.** Onboarding highlights
[Melious.ai](https://melious.ai), an OpenAI-compatible EU endpoint serving
open-weight models. You bring your own key and hold your own contract with
whoever you pick.
- Works with everything else too: OpenAI, Anthropic, Mistral, Google, Alibaba,
Nebius, OpenRouter, Ollama, or any OpenAI-compatible endpoint, including a
custom base URL pointed at your own gateway.
- **Agent templates and souls**: a template defines a responsibility, a soul
defines voice, tone bounds, coaching style, and an anti-sycophancy policy.
Both are versioned, both are improved through reviewable 1-on-1 sessions, and
both can be rolled back.
- **Grievances and 1-on-1s.** When an agent annoys you, say so — no form and no
magic phrase required. It records the grievance and brings it up in the next
1-on-1, where the two of you reconcile what changes. The result is an
assistant that evolves rather than one you configure once.
- **Contextual skills** for work that does not need a durable agent, including
the prompt generators described above.
- **Usage and impact**: tokens and requests by period, category and model for
every cloud call, plus cost, energy and CO₂e wherever the provider reports
them — currently Melious. Other providers return token counts only, so those
columns stay empty rather than being estimated.
- Turn all of it off. Lotti is useful without any inference configured.
### Sync and data
- End-to-end encrypted sync across your devices, relayed by a Synapse
homeserver whose operator never holds a key — see
[the vision](#2-end-to-end-encrypted-sync-through-a-zero-trust-relay) for how it works.
- The first device is set up against a homeserver with the included
[`tools/matrix_provisioner`](tools/matrix_provisioner) CLI, which creates
the sync account and encrypted room and produces a single-use pairing
bundle — or, in builds that offer it, by subscribing to the hosted Lotti
Sync service from the app. On Linux you can also enter an existing Matrix account directly.
- Every device after that pairs from one you already use: scan a QR code,
compare a check code shown on both screens, then complete emoji verification.
The pairing QR carries the sync account's credentials, so treat it like a
password. Until verification finishes, the new device receives ciphertext it
cannot read.
- **Your data is a file you already have.** Local SQLite with a documented
schema, plus attachments on the filesystem.
---
## Two databases, human-in-the-loop by construction
Lotti keeps *what you said and did* and *what an agent thinks* in different
databases on disk.
The **user database** is the system of record: tasks, notes, audio, photos,
time recordings, journal entries, people, events. It is treated with care, and
agents do not write into it freely.
The **agentic database** is agent working memory: definitions, memories, wake
history, reasoning traces, reports, proposals. It may grow and it may be
pruned. It can be thrown away and rebuilt without losing anything that matters
about *you*.
```mermaid
flowchart LR
subgraph Agentic["Agentic DB (prunable, recreatable)"]
A[Agent state & memories]
W[Wake cycle reports]
S[Proposals]
end
subgraph User["User DB (system of record)"]
T[Tasks & projects]
J[Journal, audio & photos]
P[People & events]
M[Metrics, habits & goals]
end
A --> W --> S
S -->|requires your approval| T
S -->|requires your approval| P
S -.->|pre-approved class only, ADR 0102| J
```
The rule matters because of how it is enforced. Agent-authored content sits in
a different file on disk and reaches the user database only through a code path
that requires your approval, or through one of the pre-approved paths
[described above](#3-memories-stay-unaltered-with-clear-provenance) and defined by
[ADR 0102](docs/adr/0102-pre-approved-agent-changes.md). Every other change an
agent wants to make to your record is a proposal you confirm or dismiss. Both
databases sync between your devices the same way.
---
## Privacy, sovereignty, and where your compute happens
Three different claims get flattened into the word "private." Lotti makes all
three. They are worth separating, because they fail in different ways.
### Your record is not collected
Lotti collects nothing — the record stays on your devices, as described in
[the vision](#1-your-memories-live-only-on-your-devices). Two things do leave,
both because you configured them: ciphertext to the homeserver you chose, and
inference requests to the provider you chose (plus, if you turn on pull request
tracking, requests to the GitHub API).
*The other side of that:* there is no server-side backup and no account
recovery, because no server holds a readable or lasting copy of your data.
Sync is the redundancy story — but it is not automatic history. Pairing a
device gives it everything written *from then on*; your existing settings and
back catalogue arrive only when you run *Send settings* and *Send message
history* from the device that already has them. Until you do, the new device is
a live peer, not a backup. Do that early, and a second device means losing one
costs you nothing.
Keep an ordinary backup as well, since replication covers hardware loss and a
backup covers everything else. Your record is a SQLite database sitting on your
own disk with a documented schema, so reading it takes a query rather than
permission. Take a consistent copy with `VACUUM INTO` while the app is running,
and bring the attachments directory along with it, because audio and images
live on the filesystem rather than in the database.
<!-- TODO: link a manual page giving the database and attachment paths per
platform. Without it, "you have access" is true and unactionable. -->
On a phone the database sits inside the app sandbox, so the practical route
there is to pair a desktop and take the copy from that peer.
Deleting works the other way round, and the order matters: delete entries *in
the app* rather than deleting the database files. Purging removes an attachment
by walking the deleted rows that still point at it, so a database removed first
leaves its audio and images orphaned on disk. Paired devices and ordinary
backups keep their own copies either way, and anything already sent to a
provider is subject to that provider's retention, not yours.
*What this does not protect against:* a compromised device. The on-device
SQLite files live inside your OS user account and are not separately encrypted
by Lotti today. That includes the sync session: its access and refresh tokens,
and its Olm identity and Megolm session keys, which the Matrix SDK stores
pickled under a key derived from the account ID rather than from a secret.
Anyone who can read your files can read your record and impersonate that
device. The sync password, AI provider keys and GitHub token are the exception:
they sit in the OS keystore. App-level at-rest encryption is a candidate for
the roadmap; until it lands, rely on your operating system's full-disk
encryption, give Lotti's on-device data the same care you would give any
personal app on the same machine, and weigh that before putting your most
sensitive categories in it.
### Your sync relay does not need your trust
Sync runs over Matrix with Vodozemac for end-to-end encryption, through a
Synapse homeserver you can run yourself or rent from anyone. Its operator does
not need to be trusted with your data: the relay only ever handles ciphertext,
keys go only to devices you verified, and events from anything else are
dropped. Both databases sync this way. The details are in
[the vision](#2-end-to-end-encrypted-sync-through-a-zero-trust-relay).
*What this does not protect against:* metadata and availability. Whoever runs
the homeserver can observe your account, which devices synced and under what
device names, from which IP addresses, when, and roughly how much. Not what.
They can also stop relaying, and how long the relay keeps ciphertext depends on
whether its operator enables retention.
### Your inference is a routing decision
Lotti has no inference backend. Every AI call goes to a provider you
configured, under your own account and API key, and you choose per category
which provider that is. Work can go to one provider while your journal goes
to another, or to none at all. That granularity is the entire point.
**Prefer zero data retention.** The provider you route a category to sees what
that category sends it. The safest cloud choice is a provider whose terms
commit to zero data retention — no prompts or responses kept after the request,
no training on them. Many providers offer that only on request or on specific
plans, so read the terms rather than the marketing page.
**Onboarding highlights a European route.** [Melious.ai](https://melious.ai) is
a German company routing open-weight models across a network of EU
infrastructure providers. They state that requests are not used for training
and that data stays under European jurisdiction. Those are their claims, on
their terms, and worth reading in full before you rely on them. Directories
such as [staysin.eu](https://staysin.eu/api.melious.ai) list Melious, but they
are compiled from providers' own published material, not from an audit.
Mistral, Google, Alibaba, OpenAI, Anthropic, OpenRouter, and any other
OpenAI-compatible endpoint work equally well. Lotti does not rank them, does
not summarise their terms, and cannot verify anyone's claims about jurisdiction,
retention, or training. Picking a provider is your decision and the due
diligence is yours.
What Lotti does instead is refuse to let the decision be invisible:
- Routing is configured per category, ahead of time, rather than negotiated in
the moment — so which provider a given piece of work goes to is a setting you
chose, not a prompt you clicked past. The flip side is that there is no
just-in-time notice at the point an agent or a transcription fires; the
onboarding and settings screens are where that decision is made and shown.
- A friendly display name is not the security boundary. The provider detail
view shows the actual base URL, the models attached to it, and every profile
that depends on them.
- Usage & Impact logs every request that left the machine and which model served
it, with token counts throughout and cost, energy and CO₂e wherever the
provider reports them.
A LAN endpoint is worth the same scrutiny. "Local" means it did not go to a
hosted provider, not that no network hop occurred.
*What this does not protect against:* a provider that does not do what its
terms say. Nothing in Lotti can verify that, and neither can anyone else from
the outside. What Lotti can do is show you every request it made, where it
went, and what it cost.
### Local models and the sweet spot for agents
Some of Lotti runs well on your own machine today. Speech is fully offline via
Whisper or Voxtral, and a local inference server handles lighter jobs. Lotti
connects to:
- [Ollama](https://ollama.com) and [oMLX](https://github.com/jundot/omlx) (MLX
on Apple Silicon), which have their own provider types in Lotti's settings;
- anything else that serves an OpenAI-compatible API, through the
*OpenAI Compatible* provider: Python-based servers such as
[vLLM](https://github.com/vllm-project/vllm),
[SGLang](https://github.com/sgl-project/sglang) and
[mlx-lm](https://github.com/ml-explore/mlx-lm), as well as
[llama.cpp](https://github.com/ggml-org/llama.cpp)'s `llama-server` and
[LM Studio](https://lmstudio.ai).
The agents are a different matter. Their sweet spot right now is models of the
size and capability of **DeepSeek V4.1 Flash** and **GLM 5.3 Flash**, both of
which perform particularly well on Lotti's agent workloads. The models people
typically run locally are not that far behind on correctness, but they are slow:
Qwen 3.8 27B, for example, lands only a few points below them, yet is far slower
even on cloud inference, occasionally stalls, and misses the day planner's
30-second budget ([the evaluation](docs/evaluations/lotti-gym-2026-09-21.md)).
On a personal machine that gap only widens: even an M4 Max with 128 GB of RAM
does not yet run the agents fast enough to keep up with daily use. Personal
hardware is improving quickly, and closing that gap is the long-term direction.
**LottiGym is how new models are judged.** Models enter the market every few
weeks, and a benchmark score says little about whether one can run a task
agent. [LottiGym](knowledge/features/ai/model-evaluation.md) runs a candidate
through Lotti's real agent harnesses — task, goal, relationship, Daily OS
planning and query exercises, several samples each — grades the results, and
records correctness, cost and latency side by side
([recent results](docs/evaluations/)). A new model earns a recommendation by
passing it:
```sh
python3 tool/lotti_gym.py assess --model MODEL_ID
```
**The practical answer today is the cloud, with zero data retention:** run
the agents on a sweet-spot-class model from a provider whose terms commit to
keeping nothing, and keep speech fully offline with Whisper or Voxtral if you
like. Image generation has no local path yet — cover art goes through Gemini or
Alibaba. As personal hardware gets faster, models of that class are expected to
move onto your own devices, until a cloud provider is a choice rather than a
necessity.
**Input on running Lotti locally is very welcome.** If you have a model,
quantisation, runtime or hardware setup that keeps up with the agents, please
share it in a [GitHub issue](https://github.com/matthiasn/lotti/issues) or on
[Discord](https://discord.gg/uuSaa8NpY).
### Energy is a routing decision too
Inference runs in a physical building on a specific grid. The **Usage & Impact**
view reports tokens and requests for every cloud call, broken down by category
and by model, so "what did my thinking cost" has an answer rather than a vibe.
How complete that answer is depends on the provider. Cost, energy, CO₂e and
water are recorded when the provider returns them with the response, which today
means Melious; every other provider reports token counts only, and those columns
stay empty rather than being filled with an estimate. That is a real limit on the
dashboard: route your work through a provider that discloses nothing, and the
impact of that work is not something Lotti can show you.
The route highlighted in onboarding publishes power usage effectiveness per
datacenter and carries Green Web Foundation verification, which is more than
most disclose. Running a model locally moves the cost onto your own hardware
and your own grid, which is a different trade rather than a free one, and the
dashboard does not measure it.
Choosing a renewable-powered route avoids outsourcing the health and climate
cost of fossil-powered compute, gas turbine generation in particular, to
communities with less power to refuse it. That is a reason to care where your
tokens get processed even if you do not care who reads them.
### If none of that satisfies you
Turn AI off. Configure no provider at all, or point every category at Ollama
with local Whisper for speech. Lotti is a complete task manager, time tracker,
and journal without a single inference call leaving the machine.
---
## In development
Built and documented, but not enabled in a default build. Turn it on under
Settings → Sections, and expect it to change.
- **Daily OS** turns a spoken or typed check-in into a plan for one day. It
reconciles what you said against your real tasks, drafts an agenda against
your actual capacity, and presents every proposed change with its old time,
its new time, and the planner's reason. Nothing is committed until you hold
the confirm control. Wrap-up at the end of the day separates what you
finished from what carries forward.
[Documentation](https://matthiasn.github.io/lotti/manual/development/plan-and-capture/daily-os)
Also under way, with no switch to flip yet: **record provenance**, so that
every entry is signed by its author and the record can prove who wrote what —
see [the vision](#3-memories-stay-unaltered-with-clear-provenance).
Next agent types on the roadmap: week planners, long-term commitment monitors,
and effort-against-goals balancers — same building blocks, different jobs.
Designed but not yet built: learning quizzes. See the
[roadmap](https://matthiasn.github.io/lotti/manual/development/roadmap).
---
## Pricing and sustainability
**Linux is and will always be free, with maximum functionality.** No paywalls,
no upsells. The same promise extends to any future fully open-source mobile
operating system. The one thing it does not include is hosted sync
infrastructure — running a homeserver is a real cost, so you either self-host
Matrix or bring your own.
**On other platforms, the basics stay free too.** Task and journal capture, the
task agent, voice transcription, the everyday loop.
**More advanced agent features may eventually be in-app purchases on platforms
where IAP is the norm** (iOS/macOS, Android, Windows). The candidates: day- and
week-planner agents, overarching project-management agents, longer-horizon
commitment monitors. The exact split is not set yet.
---
## Documentation
- **[Manual](https://matthiasn.github.io/lotti/manual/development/)** — the
full guide, available in 11 languages. Start with
[the mental model](https://matthiasn.github.io/lotti/manual/development/getting-started/mental-model)
if you want to understand the shape of the thing before installing it.
- [Connect an AI provider](https://matthiasn.github.io/lotti/manual/development/ai-and-automation/provider-setup/)
— local Ollama or cloud Gemini, and the models and profiles around them
- [Task management and voice capture](https://matthiasn.github.io/lotti/manual/development/getting-started/first-task/)
— the everyday voice-to-checklist workflow
- [Architecture](docs/ARCHITECTURE.md) — the two databases, vector clock sync,
on-device inference
- [Knowledge bundle](knowledge/index.md) — how the app actually works at
runtime, subsystem by subsystem, written for contributors and coding agents
alike; start with [security and privacy](knowledge/architecture/security-and-privacy.md)
and [sync](knowledge/features/sync/overview.md) for the details behind this
page
- [Background](docs/BACKGROUND.md) — why this exists
- [Privacy policy](PRIVACY.md) · [Security policy](SECURITY.md)
- [Roadmap](https://matthiasn.github.io/lotti/manual/development/roadmap)
Building it yourself: install Flutter via [FVM](https://fvm.app/) (the repo
includes `.fvmrc`), then `make deps`, `make analyze`, `make test`,
`fvm flutter run -d <device>`. Full setup, including the Linux audio-codec and
emoji-font packages, is in [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md).
---
## Status
The application is in active daily use and the agentic layer is real, working,
and shipping. Development happens in the open, and the [changelog](CHANGELOG.md)
is the honest version of what changed.
The agentic layer moves fastest: soul and template ergonomics, grievance
handling, and pruning strategies are all under active development, and feedback
there is especially useful. Local image generation, at-rest database
encryption, and signed provenance do not exist yet.
The manual, including every screenshot in all 11 languages, is generated from a
deterministic fixture workspace, so the documentation cannot quietly drift from
the build it documents.
**Stack**: Flutter and Dart across five platforms, local SQLite via Drift with
ObjectBox for embeddings, Whisper and Voxtral for on-device speech recognition,
Matrix and Vodozemac (Olm/Megolm) for the encrypted sync transport, Glados for
property-based tests, TLA+ and TLC for the formally verified sync protocol and
agent runtime.
---
## Contributing
The easiest way in is to say hi on [Discord](https://discord.gg/uuSaa8NpY).
Beyond that, two things genuinely help, and one thing to know up front.
**Welcome:**
- **Issues and bug reports** — the best place to start. Tell me what broke and
how to reproduce it.
- **Translations** — new languages and corrections to existing ones.
AI-assisted translation is fine **provided you contribute from a real-name,
established GitHub profile**. Translation PRs are the one kind that gets
merged.
**Not accepted: code pull requests.** Review capacity is the binding
constraint, Lotti holds people's personal data on their own devices, and
AI-generated code now arrives far faster than it can be reviewed carefully. An
unsolicited code PR will be closed unmerged, with thanks and without review —
so please open an issue instead of writing the patch. It is more useful to me
and cheaper for you.
[CONTRIBUTING.md](CONTRIBUTING.md) has the details, including what makes a
translation PR mergeable.
## License
GPL-3.0. See [LICENSE](LICENSE).
## Acknowledgments
Thanks to the [Flutter](https://flutter.dev) team, the
[Qwen](https://github.com/QwenLM) and [Mistral](https://mistral.ai) teams for
their open-weights models, [OpenAI](https://openai.com) for the
[Whisper](https://github.com/openai/whisper) weights, the
[Ollama](https://ollama.com) project, the [Matrix.org](https://matrix.org)
community, the [Vodozemac](https://github.com/matrix-org/vodozemac) authors, and
everyone contributing translations, issues, and ideas.
---
Built in public. [GitHub](https://github.com/matthiasn/lotti) ·
[Substack](https://matthiasnehlsen.substack.com)