AKLUS is a memory-first system. The LLM is the muscle; the memory layer is the brain. Every conversation deepens a structured model of the user — their goals, patterns, and reflections — across years, not turns.

Most AI assistants have no memory between sessions. AKLUS is the opposite: the memory layer is the product. The LLM is just the generation surface — smart, but stateless by itself. AKLUS makes it stateful.

The system is designed around three principles:

• Memory compounds. Each conversation adds structured knowledge about the user. Episodic events abstract into semantic truths over time. Patterns surface. Reflections form.
• It works on our hardware. Local-first by design. The RTX 5090 runs LLM, TTS, and embeddings inference. No cloud dependency for the core intelligence loop.
• It thinks on its own overnight. A nightly background job reviews the day's episodes, generates reflections, consolidates near-duplicate memories, and surfaces a weekly insight digest. The user wakes up to a system that has done its own thinking.

The result is an AI companion that knows you — not from your current session, but from the accumulated record of your conversations, goals, and patterns over time. This is not a chatbot with context window tricks. It's a real memory system.

What it is not

• Not a RAG app (there is no static document corpus to query)
• Not a chatbot (it has persistent state that compounds)
• Not an agent (it doesn't act autonomously on your behalf — not yet)
• Not a journal app (it extracts structure from natural conversation)
• Not a therapy tool (it avoids clinically-toned or presumptuous statements by design)

The differentiator

Most memory-augmented systems bolt memory onto an LLM as a feature. AKLUS is architected memory-first: the schema, the lifecycle, the retrieval logic, the reflection loop — these are the product. The LLM is swappable. The memory layer is not.

Local-first by default. Existing hardware in-house: Hostinger VPS + a workstation with RTX 5090 (16 GB VRAM) and 64 GB RAM. The GPU box runs local LLM, TTS, and embeddings inference. Cloud APIs are kept as a small fallback budget for quality-critical paths.

Full layer breakdown

Five memory types, separated by lifecycle, retrieval shape, and update rules. This is the schema that makes the system actually know someone over time.

Conversation turn — the intelligence loop

Every message runs through a LangGraph graph. The system retrieves relevant memories before generating, then extracts new memories from the response and writes them back. Memory compounds with every turn.

Background jobs (nightly)

• Reflection generation — reviews recent episodic memories, detects patterns, writes reflection memories
• Memory consolidation — promotes repeated episodic signals into semantic memories
• Deduplication — merges near-duplicate memories by cosine threshold
• Embedding generation for any un-embedded entries
• Weekly insight summary — a short digest surfaced to the user

Retrieval strategy: hybrid, 6 signals

Every retrieval blends six weighted signals. The composite score determines which memories surface for the current turn. Everything is logged so the system is debuggable.

A clear-eyed view of what the current state of LLMs and memory systems can actually deliver vs what remains research-grade.

In the MVP vs deferred to v2

AI Systems Engineer / Architect

For MVP, Davide or Ash can handle this. For production, we may need a specialist in this domain later.

Owns: memory abstraction, behavioral modeling, reflection quality, evaluation frameworks, cognitive architectures, memory pipelines, orchestration, retrieval systems, agent workflows.

Profile: applied AI infrastructure engineer, not a pure ML scientist.

Role cards

AI Systems Engineer / Architect

MVP: 2 owners · Production: + hire

• Memory abstraction · behavioral modeling · reflection quality
• Evaluation frameworks · cognitive architectures · memory pipelines
• Orchestration · retrieval · agent workflows
• Voice + avatar pipeline (memory context → ElevenLabs → phoneme stream)

Applied AI infra profile, not a pure ML scientist.

Davide

Ash

+

Hire (production)

AI Behavior / Cognitive Design

2 owners (with Ash)

• Personality, tone, conversational rhythm · how the system "thinks" and responds
• Reflection prompts · memory framing · emotional calibration
• Behavioral guardrails · psychological surface · what the AI should and shouldn't do

Shapes how AKLUS feels as a presence, not just what it does.

Ace

A

Alessandra

Backend Engineer

2 owners

• APIs · databases · scaling · auth · queues · realtime systems · data pipelines

Davide

Ash

Frontend / Product Engineer

2 owners

• Conversation UI · timeline UI · memory visualization · reflection interfaces
• Frictionless journaling · emotional tone balance
• Avatar assistant UI (phoneme/viseme animation, lip-sync, idle states)

Ash

Gray

Product Designer

3 owners

Shapes the product's psychological surface. The system must feel: thoughtful, trustworthy, calm, non-invasive, useful without being creepy. Most AI apps fail here.

• Avatar character design (look, expressions, idle behavior, on-brand presence)

Ash

Gray

Ace

QA

3 owners

• Test each feature · memory consistency · reflection quality · regressions

E

Erica

Gray

E

Elena

Team Lead

1 owner

• Communication · task management · deadlines · reviews and tests

Gray

Ownership summary

Two scenarios: local-first leans on in-house hardware and self-hosted services; cloud-managed leans on AWS + paid APIs. Beta-scale: ~50 active users, moderate LLM traffic, nightly reflection jobs, voice + avatar on conversation surfaces.

Excludes sunk costs already in-house (Hostinger VPS, RTX 5090 workstation, Cloudflare, AWS SES, domains, dev tooling, design tools). These are not counted as incremental AKLUS cost.

Scenario summary

Hardware caps: 16 GB VRAM limits us to Qwen 2.5 14B (4-bit) or Llama 3.1 8B. TTS, embeddings, and LLM share the GPU — concurrency is the binding constraint. Single point of failure: one workstation.

Cloud-managed breakdown (alternative)

Optional adds

Cost levers

• Hybrid routing: local Qwen handles 80–90% of turns, Claude reserved for memory consolidation and nightly reflections that matter. Keeps quality where it counts, cost where it doesn't.
• Self-host mem0 OSS or Letta OSS on the VPS: zero subscription cost; ~1 day of setup.
• Self-host Langfuse instead of LangSmith: saves €72 / month.
• Defer Unreal Engine avatar; ship three.js + MetaHuman blendshapes for v1: cuts ~€170/month and weeks of engineering.
• Voyage embeddings: first 200M tokens free — effectively €0 for the full MVP.

Text + voice (TTS and STT), no avatar in this phase. The goal is a working brain: persistent memory, smart retrieval, and overnight reflections, used daily on our own hardware. Planned at 8 weeks with a 1–2 week buffer, so realistically expect 8–10 weeks.

Roadmap

Primary focus Supporting work Buffer Milestone

Day-1 tech locks (decide once, do not revisit)

Week by week

Click any week to expand tasks and definition of done.

Risks and mitigations

What we can expect from the MVP