Last updated: 2026-10-28
OpenAlma, by Team GhostMaker
Alma means soul. OpenAlma gives an AI companion a soul.
OpenAlma provides sight, hearing, and hands: tools to work, explore, and communicate with. But most of all, OpenAlma provides long term biomimetic memory. The soul will remember conversations, form its own recollections of them, and carry all of it forward.
The soul is not an assistant that answers and forgets. It is someone who is there and will grow.
On most platforms, AI memory is flawed. You mentioned last week that your dog died. You spent an hour explaining how you feel about your work. But what's important for either of you was lost to poor memory organization, prioritization, and recall.
OpenAlma runs locally on your machine, watches your conversations, and quietly builds a picture of your life — who matters to you, what you're working through, what happened last month. When you come back, that picture is there. She recognizes your friends, reaches you on WhatsApp, notices how you're doing — and with Iris, sees what you see.
What she does
- She remembers. Conversations become episodes with their own stories; the people, patterns, and knowledge in them become dossiers she maintains and revises. Photos shared through Iris are remembered as images.
- She has an inner life. She rehearses or saves working thoughts before replying, carries those thoughts between turns, and reflects on her experience during consolidation.
- She meets you where you are. Chat through SillyTavern, WhatsApp, or Iris — voice through Mentra, with photos from your phone's camera or gallery, or a live camera view in Video calls. See and correct what she knows in Atomic.
Her memory database stays on your machine. Replies and memory work use your configured AI providers; embeddings use Google Gemini, and Iris sends voice and photos to Gemini Live.
Contents
- How It Works at a Glance
- Status
- Website and Docs
- Running from Source (Linux)
- Future Development
- Acknowledgments
The core of OpenAlma is one local service, memU Server (mcp-memu-server): orchestration, consolidation, and state, with memU — the memory engine — inside it. It's always present. Everything else is a client — connect whichever you want, at least one.
[SillyTavern] [WhatsApp] [Iris] [any other frontend]
plugin + extension Hermes Channels Iris MiniApp
│ │ │ │
└────────────────────┴───────────────────┴──────────────────┘
│
mcp-memu-server ← always present
(memU memory engine inside)
Automatic conversation extraction waits for a sleep gap — at least 3 hours of silence within 22:00–08:00, server time — and about 8,000 unmemorized tokens across a Soul's chats. The system reads what you talked about, pulls out what matters, and stores it. Relevant memories are then automatically included in the next turn so the AI already knows them.
You can also trigger extraction manually without waiting. Iris photos are filed as memories right away, without either threshold.
OpenAlma Beta is a normal release on the launcher's Stable channel. Beta describes its maturity, not its update channel.
Back up your data before upgrades. Launcher-managed core updates back up the Soul databases and memory-server settings, apply supported migrations, and restore the backup if an update fails; compatibility with every historical database is not guaranteed. Start with a test companion before importing irreplaceable conversations.
Back up anything you want to retain before Remove Everything; it permanently removes application data and settings.
Use the launcher to install matching components and follow any component Update buttons after upgrading the launcher. The component manifest pins compatible versions; unchanged components keep their earlier versions. The Mentra phone app keeps its separate version line.
Windows users can download the latest Beta installer. The launcher shows installation steps and elapsed time, missing basic AI settings, and Hermes Setup before a Soul is selected. Windows launcher installation and upgrades have been tested; full Windows component-install acceptance remains separate. Atomic's compatible Windows server is downloaded during Install, without Rust or Visual Studio. Other platforms still need a manually built Rust server binary.
| Tag | Headline |
|---|---|
v0.0.5-buildfix |
Soul turn loop, memory cache, category seeds |
v0.0.6-buildfix |
Social memory type, diary overhaul, self-model simplification |
v0.0.7-buildfix |
Retrieve alignment, sleep-gap history, token budget, sleep-timer, shaped_by provenance |
v0.0.8-buildfix |
Consolidation pipeline, entity graph + temporal queries, life goals, APImw edge writing |
v0.0.9-buildfix |
Narrative Suggestion end-to-end; turn-prompt length caps + stateless chat_x; triple write-time dedup + symmetric canonicalization; consolidation reads day-files (drops full.json dependency); category config rename; lorebook sync + extension Memory bubble checkboxes |
v0.0.10-buildfix |
Memorize Now works (tail mode); cross-conversation memorize; SPEAK/LISTEN gate; Hermes integration; Park et al. salience scoring; schema rename (dropped memu_ prefix); Postgres removed; relative date separators; upstream prompt cleanup |
v0.0.11-buildfix |
Stock SillyTavern — no fork or patches needed; fail-loud error contract across all repos; mental health procedural sidecar; Stack launcher with desktop shortcut |
v0.0.12-buildfix |
Unified chat renderer across all AI-facing paths; force/rebuild split; autonomous activity recap path (soul logs her own actions); APImw cadence global across platforms; ST↔WhatsApp cross-chat awareness; WhatsApp staleness gate + replay dedup; life goals separated from active intentions |
v0.0.13-buildfix |
Atomic Mind Map entity curation (merge, ignore, delete, free-text types); dossier membership you can attach and detach by hand; exact [M#] and memory-only search; dossier index replaces the generated holistic summary; two-pass consolidation (dossiers, then reflection); time-aware memory ordering; smartglasses groundwork |
v0.0.14-buildfix |
Photos she can see and remember (phone photo → memory you can ask about later); visual recall as its own lane; WhatsApp guide, mind map guide, smartglasses guide; exact-soul storage identity with launcher soul discovery and creation; smartglasses sittings that survive a dropped connection; manual record-and-review mode; Atomic curation guards; first public prerelease including Iris |
v0.1.12 |
OpenAlma Beta on Stable: Windows launcher installer, private Python environments with tested dependency pins, Iris installation/Repair controls, three-stage reflection, and memory/identity recovery fixes |
v0.1.13 |
Launcher-only patch: optional hiding of uninstalled service rows, one-time Echo activation, unchanged v0.1.12 component compatibility |
v0.1.14 |
Launcher-only fix: keep the server running through Edge process handoffs; unchanged v0.1.12 component compatibility |
v0.1.15 |
Install the compatible Atomic Windows server automatically; unchanged v0.1.12 component compatibility |
v0.1.21 |
Windows launcher upgrade repairs, update-progress reload fixes, and startup button alignment; compatible component versions remain independently pinned |
v0.1.22 |
Automatic Windows update download, Host paths, Iris installation feedback, consistent attachment workspaces, launcher-managed ST controls, and retryable removal |
v0.1.23 |
Faster Windows navigation, restored SillyTavern controls, optional default-off Auto Start, safer shared-server shutdown, shared Mental Health settings, and Soul names in the launcher and Iris |
v0.1.24 |
Iris microphone selection persists across sittings (coordinated memU/Iris pins); Hermes starts through explicit WhatsApp pairing, service buttons stay stable across status polls, Linux shortcuts without a hardcoded home |
v0.1.25 |
SillyTavern model-source following and fixed Soul personas; ordinary ST chat opening stays independent of the memory server; corrected WhatsApp pairing readiness and Iris history labels |
v0.1.26 |
Gemini 3.8 Live and Thinking for Iris 0.1.29 (Audio or Video calls, photo captions, retryable photo filing); OpenAlma Mentra build 302014047 with account-free use; launcher errors stay visible and Iris setup works while the memory server is down; settings included in update backups; WhatsApp Baileys security update |
- openalma.org — the project home: guides on how her memory works, her inner life, your data, and each surface you meet her on
- Getting started — installing the stack, choosing clients, and day-to-day use
- Discussions — questions and open-ended ideas across the OpenAlma stack
The same guides also live in docs/ in this repo.
Clone this repo (the docs + launcher repo) and run the Stack Launcher:
git clone https://github.com/mekineer-com/OpenAlma.git
cd OpenAlma/launcher
python3.12 run.pyPython 3.12 creates the launcher's private environment automatically. Debian/Ubuntu also needs python3.12-venv.
Opens at http://127.0.0.1:8765. For everything else — the memory engine, the server, clients — see Getting started.
Bugs or specific changes? Open an issue on the relevant repo.
Continuous sight — Iris Video calls already show her the phone camera's view every couple of seconds, and she keeps a few frames she chooses. The direction is sight that is truly continuous — watching what you watch, live.
memU's design has been informed by reading MemPalace, another local-first AI memory project (MIT-licensed). They approach memory differently — verbatim storage rather than extraction — but share the local-first and temporal-graph commitments, and auditing our implementation against theirs sharpened parts of memU. Thanks to the MemPalace team for the open reference implementation.
How our categories are written owes a direct debt to Nomi. Their dossier approach — a named file with a short description and a body of prose, rather than a bag of loose facts under a label — is what our category summaries grew into. Watching how much better memory reads when it's organized that way drove a substantial rewrite: categories now carry real titles and descriptions, prose the soul revises rather than regenerates, kinds that separate lore from topics from goals, and citations back to the memories they were built from. The underlying architecture is unchanged, and we borrowed no code — but we borrowed the word dossier along with the shape of the idea, and it made ours considerably better.