GamiWays
01·Reusable foundation

GamiWays Project

Collaboration between Memoways (Geneva, 14 years of interactive video expertise) and Gamilab (voice-first AI startup, Audiogami SDK).

GamiWays Research Portal·Gamilab × Memoways·Geneva, Switzerland

Technical challenges

The problems the Core must solve across projects.

These challenges are not an abstract promise. They come from the concrete fields of Où est Ava ? and Edugami, then GamiWays addresses them with reusable rules and measurements.

Make the exchange genuinely live

Reduce the perceived wait between an intention, a voice response and a video presence without promising fluidity that has not been measured.

Why this matters for the Core +

The Core must observe delays at every step, choose graceful degradation when needed and preserve continuity of the experience instead of optimising a single isolated metric.

Keep just enough context

Resume a conversation without turning every session into an infinite, costly and unreadable history.

Why this matters for the Core +

The question is not to remember everything. It is to distinguish what serves the current turn, what documents progress and what belongs durably to a relationship with a person.

Direct the experience, not only generate a response

Keep scenario rules, character role, available knowledge and conditions for moving between steps together.

Why this matters for the Core +

The prototypes reveal narrative and learning choices that should not be left to one prompt. The Core must make them configurable, inspectable and testable.

Remain deployable and reversible

Be able to integrate a commercial layer when it accelerates a field deployment without locking the experience, data or product decisions into it.

Why this matters for the Core +

Interface contracts, observability and exit tests matter as much as choosing a provider. The Core stabilises these conventions so they are not reinvented for every project.

Diagrams & reference points

The technical trade-offs to make visible before optimising.

The diagrams are open to provide an overview. You can collapse them individually after reading.

Conversation flow and measurement points

This diagram follows a response from speech to avatar. It shows where time is spent and which points must be measured before optimising.

SEQUENCE DIAGRAM — CONVERSATIONAL EXCHANGE (TARGET <2s)Hover an arrow for detailsUserASR/STTOrchestratorMemoryLLMTTSAvatarSpeech (audio)t=0Transcribed text+300msContext query+350msContext + profile+400msEnriched prompt+420msResponse (stream)+900msText → synthesis+950msMemory update+960msAudio + phonemes+1100msVideo + audio sync ⚠+1500ms ✓

Hover a diagram arrow to see latency budget and each component's role.

Target timings · TTS+Avatar parallelization possible · LLM → TTS streaming

Role orchestration

This diagram clarifies how experience roles, journey rules and services intervene during a session. It prepares the conventions the Core can reuse.

FREEDOM DEGREE — PER CONVERSATION NODEEach node defines its own deterministic ↔ organic balance0%ScriptedFixed sequenceNo AI generation→30%GuidedMandatory contentAI rephrases only→50%BalancedHard + soft constraintsAI adapts to user→70%CreativeFew constraintsAI drives dialogue→90%+OpenTopic boundaries onlyFull conversational AIPEDAGOGICAL MODE0–50% · Mandatory content coverage · Strong pedagogical controlAI adapts tone, not contentNARRATIVE MODE50–90%+ · Topic boundaries only · Character personalityAI drives narrative evolutionR&D challenge: guarantee mandatory content coverage (deterministic) while maintaining conversational naturalness (organic)Hypothesis H4 — Axis 3

Target architecture

What GamiWays chooses to make reusable.

The Core does not aim to make Où est Ava ? and Edugami uniform. It stabilises the conventions that prevent rebuilding context, memory, rules and observation for every new experience.

01

Experience interfaces

Journeys, characters and production tools use clear contracts instead of logic specific to each screen.

02

Journey orchestration

The engine manages session state, progression rules, the available avatar and events to observe.

03

Context, memory and knowledge

Context is assembled in bounded form so the experience remains coherent without becoming opaque or too slow.

04

Adapters and operations

Model, transport, persistence and observability services remain replaceable behind explicit operating conventions.

Core diagrams

The conventions the foundation makes reusable.

These two reference points connect architecture layers and memories to what an experience must retain, observe and resume.

GamiWays Core target architecture

This diagram connects experience, orchestration, context and operations layers. It shows what the Core aims to stabilise across Où est Ava ?, Edugami and future experiences.

delivered in repositoryexplicit boundary
Currently documented architectureEvery block has a tooltip; arrows describe exchanges, not a latency promise.CONSUMER APPLICATIONSPlayer webscénarios · chat · voix optionnelleAuthoring adminscénarios · avatars · sourcesOperator consolediagnostic local · inspectionEvaluation CLIcomparaisons scriptéesexternal HTTP consumerHTTP authentifiéCORE API · FASTIFY/v1 JSON · SSE · bounded audiox-api-key · shared contractsCORE RUNTIMEConversationAvatar · session · SSE turnsasynchronous Game MasterContext & knowledgebounded memory · ingestionpgvector + traceable selectionADAPTERSResponse modelsavatar / GM / memory rolesOptional voiceDeepgram STTINFRASTRUCTUREDataHealth & metricsOPEN PRIORITIESguided progression · hybrid responsereal Où est Ava ? validation · recovery

Source: docs/ARCHITECTURE.md, TECH_STACK.md, PROJECT_STATUS.md and EPICS.md · repository reviewed 10 October 2026

↔ Zoom
Click to expand
Retained memory architecture

This diagram distinguishes turn, session and relationship memories. It helps decide what remains temporary, should be measured or becomes useful continuity for the person.

CONVERSATIONAL MEMORY ARCHITECTURE — 3 LAYERSLLM / AgentOrchestratorL1Node MemoryShort-termCurrent conversationNode variablesEmotional stateLLM Context WindowCost: HIGHL2Session MemoryMedium-termVisited node pathProgression scoreSummarized historyVector DB / RAGCost: MEDIUML3User MemoryLong-termLearning profileHistorical sessionsCross-session patternsPostgreSQL + SLMCost: LOWGoal: -90% context window tokens · +26% accuracy (Mem0, 2025)
01
Infrastructure & Tech Stack

Operational today vs. R&D required.

Up-to-date synthesis of GamiWays Core (shipped epics) and the Où est Ava ? prototype — against remaining R&D work.

Actual state of GamiWays Core (shipped epics) and the public Où est Ava ? prototype — vs. remaining R&D work.

GamiWays Core

Shipped

Context Engine v2

Deterministic 7-dimension assembler — memory, scenario, RAG, GM directives, persona — with token budget and selection trace.

Context Engine v2 replaces ad hoc prompt assembly: seven dimensions (short-term memory, working memory, long-term facts, scenario config, RAG, GM directives, persona) merge under a strict token budget. Every trim is traced — essential for debugging avatar responses in production and comparing Où est Ava ? vs Core runs.

  • →EPIC 5.2 shipped (May 2026)
  • →Precedence policy + trimming telemetry
  • →Injected on every avatar and GM turn

Où est Ava ? prototype

Remaining R&D

Tap an item to show its detail below.

02
Core Engine — Vision & Principles

From content delivery to interactive experience orchestration.

The GamiWays Core is a headless orchestration engine for guided interactive experiences. Explore the content → orchestration shift, then domains and principles — with links to portal details.

The shift

The Game Master drives the experience — the Core assembles context.

Durable sessions, layered memory, typed RAG, GM directives, unlockable avatars and runtime traces: this foundation is delivered. A standalone guided-progression engine and real-scenario validation remain explicitly open.

Select a domain — prototype example, key principles and portal pages to go deeper.

📚

Learning

Adaptive journeys with objectives and remembered progression.

The avatar guides learners through structured objectives: the GM unlocks steps, injects hints, adapts pace. Episodic memory avoids starting from scratch each session.

Concrete example

Edugami / Dilemme Plastique prototype — plastic clues, Race for Water.

Related principles

01 · Experience First03 · Context is the Product06 · Measure Everything That Matters

Tap an item to show its detail below.

03
Key Concepts

Avatar + Game Master: two agents, one experience.

The Core is structured around precise concepts — each entry details the role, APIs, and what Où est Ava ? does with it in practice.

Shared Gamilab × Memoways vocabulary — from the Avatar + Game Master duo to runtime building blocks.

The two agents

Avatar

Conversational actor

Interaction surface — persona, own memory, voice and face. Does not decide the global narrative arc.

The avatar assembles a prompt from the Context Engine: editorial identity (7 fields in Où est Ava ?), user facts, visible RAG chunks, active GM directives. It responds in streaming; the GM works in parallel.

  • ·POST /v1/conversations/{id}/messages
  • ·Per-avatar model override
  • ·visibleToAvatarIds RAG filtering

Où est Ava ?

In Où est Ava ?: Max, Où est Ava ? and other film characters — each RAG chunk is character_id-scoped to prevent persona leakage.

Experience model

Context & memory

Tap an item to show its detail below.

04
Build Roadmap

Three phases, one vision.

Click a phase to explore goals and epics — data synced from the development repository.

Select a group to distinguish delivered capabilities from open priorities.

GROUP 01Reviewed 10 October 2026

Delivered conversational runtime

A working headless runtime with persistent sessions, asynchronous orchestration, bounded context, operational authoring and a public player surface.

Delivered foundation · open priorities· repository state reviewed 10 October 2026

Items · 5/5 delivered (100%)

2.xAvatar, sessions and streaming turns

Persistent users, scenarios, avatars, sessions, conversations and messages; direct synchronous and SSE turns with interruption-safe persistence.

4.xAsynchronous Game Master

Post-turn planning, routing proposals, runtime events and guarded decisions without making the avatar reply wait.

4.2–5.2Memory, knowledge and context

Bounded working, episodic and user-fact memory; typed knowledge ingestion, profile-aware pgvector retrieval and deterministic context assembly.

6.1–7.1Authoring and player applications

Admin scenario builder, read-oriented console and public web player with scenario discovery, SSE updates and text fallback.

8.6Scripted conversation evaluation

An authenticated external CLI compares model configurations with semantic judging and runtime latency, token and cost reports.

Opportunities & Technical Gaps

Analysis of identified gaps in the state of the art and their translation into product and business opportunities for Gamilab and Memoways.

CRITERIA COVERAGE BY SOLUTION CATEGORY

Read by category

Commercial platforms

HeyGen · Tavus · Synthesia · D-ID

⚡Real-time
~PartialR&D
🎭Behavioral fidelity
✗NoR&D
🔒Data sovereignty
✗No
🧠Conversational memory
✗NoR&D
🎬Narrative control
✗No
👁️Emotional perception
~PartialR&D
🎨Multi-style avatar
✗No

Open-source models

SadTalker · Wav2Lip · MuseTalk · Simli

⚡Real-time
~PartialR&D
🎭Behavioral fidelity
~PartialR&D
🔒Data sovereignty
✓Yes
🧠Conversational memory
✗NoR&D
🎬Narrative control
✗No
👁️Emotional perception
✗NoR&D
🎨Multi-style avatar
~Partial

Academic prototypes

VASA-1 · AvatarForcing · A²-LLM

⚡Real-time
~PartialR&D
🎭Behavioral fidelity
✓YesR&D
🔒Data sovereignty
~Partial
🧠Conversational memory
~PartialR&D
🎬Narrative control
✗No
👁️Emotional perception
~PartialR&D
🎨Multi-style avatar
~Partial

LemonSlice (LS-2.1)

Dec 2025 · $10.5M YC+Matrix · 20B DiT · 20 FPS

⚡Real-time
~PartialR&D
🎭Behavioral fidelity
~PartialR&D
🔒Data sovereignty
✗No
🧠Conversational memory
✗NoR&D
🎬Narrative control
✗No
👁️Emotional perception
~PartialR&D
🎨Multi-style avatar
✓Yes

GamiWays target

Memoways × Gamilab — R&D 2025–2028

⚡Real-time
✓YesR&D
🎭Behavioral fidelity
✓YesR&D
🔒Data sovereignty
✓Yes
🧠Conversational memory
✓YesR&D
🎬Narrative control
✓Yes
👁️Emotional perception
✓YesR&D
🎨Multi-style avatar
✓Yes
✓Yes
~Partial
✗No
R&DFundamental research axis

Detailed gap analysis

Select a domain to compare gap, state of the art, and GamiWays opportunity.

Six structural gaps — and how Core + Où est Ava ?/Edugami prototypes address them.

Conversational memory

Critical

Identified gap

No production-grade solution for 1h+ sessions without token explosion — partially addressed by Core v2/v3.

Best state of the art

Mem0 (-90% tokens, +26% accuracy) — not validated for multi-session avatars.

GamiWays opportunity

Delivered 3-layer architecture + avatar-specific SLM distillation — Où est Ava ? validates in real conditions.

Tap an item to show its detail below.

05
Competitive Gap

No solution combines all 5 criteria.

Rows reflect services currently documented in the STT, TTS and video avatar comparisons. A cell qualifies a documented service capability, not a promise of a complete pipeline. Click column headers to sort.

Voice PipelineSTT · RAG · TTS

13 solutions shown

Audiogami (Gamilab)

STT · Suisse

Real-time <2s
✓ Yes
Behavioral fidelity
✗ No
Sovereignty
✓ Yes
Conv. memory
✗ No
Narrative control
✗ No
Cartesia Sonic 3.6

TTS · Cloud

Real-time <2s
✓ Yes
Behavioral fidelity
◐ Partial
Sovereignty
✗ No
Conv. memory
✗ No
Narrative control
✗ No
Deepgram Flux TTS

TTS · Conversationnel

Real-time <2s
✓ Yes
Behavioral fidelity
◐ Partial
Sovereignty
◐ Partial
Conv. memory
◐ Partial
Narrative control
✗ No
Deepgram Nova-3

STT · Cloud

Real-time <2s
✓ Yes
Behavioral fidelity
✗ No
Sovereignty
◐ Partial
Conv. memory
✗ No
Narrative control
✗ No
Eleven v4

TTS · Cloud

Real-time <2s
✓ Yes
Behavioral fidelity
✓ Yes
Sovereignty
✗ No
Conv. memory
✗ No
Narrative control
✗ No
Fish Audio Drama 3

TTS · Preview

Real-time <2s
◐ Partial
Behavioral fidelity
◐ Partial
Sovereignty
✗ No
Conv. memory
✗ No
Narrative control
◐ Partial
Fish Audio S2.1 Pro

TTS · Cloud

Real-time <2s
◐ Partial
Behavioral fidelity
◐ Partial
Sovereignty
◐ Partial
Conv. memory
✗ No
Narrative control
✗ No

GamiWays Target

Target

Real-time <2s
⚗ R&D
Behavioral fidelity
⚗ R&D
Sovereignty
✓ Yes
Conv. memory
⚗ R&D
Narrative control
✓ Yes
Grok Voice Think Fast 2.0

Speech-to-speech · Cloud

Real-time <2s
✓ Yes
Behavioral fidelity
◐ Partial
Sovereignty
✗ No
Conv. memory
✗ No
Narrative control
✗ No
Inworld Realtime TTS-2 + Flash

TTS / agent · Cloud

Real-time <2s
✓ Yes
Behavioral fidelity
✓ Yes
Sovereignty
◐ Partial
Conv. memory
◐ Partial
Narrative control
◐ Partial
Voxtral ASR (Mistral)

STT · Open weights

Real-time <2s
◐ Partial
Behavioral fidelity
✗ No
Sovereignty
✓ Yes
Conv. memory
✗ No
Narrative control
✗ No
Video AvatarsReal-time · Behavioral · Sovereign

11 solutions shown

Anam.ai Cara-3/4

Vidéo live · API

Real-time <2s
✓ Yes
Behavioral fidelity
◐ Partial
Sovereignty
◐ Partial
Conv. memory
◐ Partial
Narrative control
◐ Partial
Beyond Presence Genesis 2.0

Vidéo live · EU

Real-time <2s
✓ Yes
Behavioral fidelity
◐ Partial
Sovereignty
◐ Partial
Conv. memory
◐ Partial
Narrative control
◐ Partial
D-ID Expressive V4

Vidéo live · API

Real-time <2s
✓ Yes
Behavioral fidelity
✓ Yes
Sovereignty
✗ No
Conv. memory
◐ Partial
Narrative control
◐ Partial

GamiWays Target

Target

Real-time <2s
⚗ R&D
Behavioral fidelity
⚗ R&D
Sovereignty
✓ Yes
Conv. memory
⚗ R&D
Narrative control
✓ Yes
HeyGen LiveAvatar LITE

Vidéo live · API

Real-time <2s
✓ Yes
Behavioral fidelity
◐ Partial
Sovereignty
✗ No
Conv. memory
✗ No
Narrative control
◐ Partial
LemonSlice LS-2.1 (Self-Managed)

Vidéo live · GPU

Real-time <2s
◐ Partial
Behavioral fidelity
✓ Yes
Sovereignty
✓ Yes
Conv. memory
✗ No
Narrative control
◐ Partial
MuseTalk v1.5

Vidéo live · Open source

Real-time <2s
✓ Yes
Behavioral fidelity
◐ Partial
Sovereignty
✓ Yes
Conv. memory
✗ No
Narrative control
✗ No
Runway Characters

Vidéo live · API (5 min fixes)

Real-time <2s
◐ Partial
Behavioral fidelity
◐ Partial
Sovereignty
✗ No
Conv. memory
✗ No
Narrative control
✗ No
Simli Trinity-1

Speech-to-video · API

Real-time <2s
✓ Yes
Behavioral fidelity
◐ Partial
Sovereignty
◐ Partial
Conv. memory
✗ No
Narrative control
✗ No
Tavus Phoenix-4.5 + Sparrow-2

Vidéo live · API

Real-time <2s
✓ Yes
Behavioral fidelity
◐ Partial
Sovereignty
✗ No
Conv. memory
◐ Partial
Narrative control
◐ Partial
bitHuman Essence 2 / Expression 2

Vidéo live · Self-hosted

Real-time <2s
✓ Yes
Behavioral fidelity
✓ Yes
Sovereignty
✓ Yes
Conv. memory
◐ Partial
Narrative control
◐ Partial