Memory System Architecture Implementation Alternatives

[lucas-oma]

Clear and concise description of the problem

Please correct me if I'm wrong in any point

Looks like AIRI currently lacks a memory system for RAG. The existing DuckDB setup seems to only only creates a temporary table with no persistence or advanced memory functionality for web/app.

I intend to write a PR for this (if issue is approved)

Suggested solution / Ideas

Option 1: Standalone Memory-Core + Database Implementations

A shared TypeScript/JavaScript library (@proj-airi/memory-core) that provides database-agnostic memory operations, with platform-specific implementations for DuckDB (web) and PostgreSQL (desktop).

Pros:

• Platform Flexibility: Each platform uses its optimal database (DuckDB for web, Postgres for rust/tauri app)
• No Infrastructure: No servers needed for basic usage

Cons:

• Data Isolation: Memories don't sync between platforms
• Setup Complexity: Each platform needs database setup (code: classes, methods, etc)
• Limited Sharing: Can't share memories across devices

Technical Requirements

Memory-Core Package (@proj-airi/memory-core):

• Database-agnostic interfaces for memory operations
• Shared consolidation and categorization logic
• Common types and utilities
• Platform-agnostic embedding service interface

DuckDB WASM Persistence (for web UI):

• Configure DuckDB for IndexedDB storage
• Implement database export/import for backup
• Handle browser storage limits

PostgreSQL Integration (for rust/tauri base app):

Similar to what has been done for telegram-bot

• Use the pgvector extension
• Implement HNSW vector indexing
• Handle local PostgreSQL setup

Alternative

Option 2: Docker Memory Service

A centralized memory service running in Docker that provides a shared memory backend for all AIRI applications (web, desktop, Discord bot, Telegram bot) through a REST API.

Pros:

• Data Synchronization: Shared memories across all platforms (Web, desktop, Discord, Telegram, all share same memory if configured)
• Centralized Management: Single source of truth/memory
• Advanced Features: Full PostgreSQL + pgvector capabilities even for web app
• Cloud Hosting: Can be hosted on personal server ( or Moeru's cloud in the future!) for shared memory across devices. Could also be useful for sharing across computers, and maybe smartphones in the future.
• Consistency: All platforms use identical memory logic and features
• Future-Proof: Easy to extend with new features for all platforms simultaneously

Cons:

• Infrastructure Required: Needs server setup (user will have to run docker alongside the usual app)
• Complexity: Docker setup and maintenance

Technical Requirements

Memory Service API:

• REST API for memory operations
• WebSocket support for real-time updates

Additional context

My preference: Option 2 - Docker Memory Service

Why Option 2?

One Update, All Platforms

• Fix a memory bug? All platforms get the fix instantly
• Add new memory features? Available everywhere immediately

Shared Memory Across Devices

• Chat on web --> Continue on desktop --> Same memories
• Discord bot knows your web conversations (if you decide to connect the same memory)

Advanced DB features

• No platform restrictions: Not limited by browser storage, IndexedDB quotas, or web app constraints - can use any database technology

Multi-Persona Support (future feature?)

• Instead of one AIRI with one memory, you can create multiple AIRI "personalities" each with their own separate memories.
• Example: One for personal assistant (work focused), one for gaming, etc.

Questions for Community

1. Which option do you prefer? Standalone or Docker service?
2. Technical Priorities: What memory features are most important?

---

Tags: enhancement, memory-system, rag, architecture, database

Validations

• Follow our Code of Conduct
• Read the Contributing Guide.
• Check that there isn't already an issue that request the same feature to avoid creating a duplicate.

[fly6516]

I just happened to come across this issue. First off, I want to clarify that I have no affiliation with this project—I'm simply a highly interested user, so my perspective is purely personal. That said, I do have my own thoughts. I believe Docker storage isn't particularly well-suited for software with a Windows version. For the desktop version, pairing SQLite with relevant vector search might be a better approach?
While Docker offers cloud hosting advantages, I suspect many users prefer local hosting for privacy and security.
That said, an option to switch to Docker deployment for server owners could be convenient.
You might ask why not use WSL or Docker Desktop. My experience with both was subpar—I previously tested backend container control logic on my laptop using these systems during backend development. While Microsoft's updates have largely eliminated blue screen bugs and system corruption issues, both solutions still consume significant disk space and do impact performance to some degree.
This is just my perspective, open to discussion. Much of my understanding of RAG is theoretical, but I'm genuinely intrigued by this long-term memory-related foundation. I'm simply bringing this up for discussion since I haven't seen much engagement on this issue. If I've misunderstood anything, please feel free to correct me.

[lucas-oma]

Hi @fly6516 , I think you have a good point... I've mostly been working on Mac/Linux, so I'd forgotten about the Windows side of things.

Right now I'm working on having the core memory "server" logic as a Node module that can also be run with pnpm or npx, without requiring Docker (though personally I'll still prefer using Docker). The module just needs a connection URL to a PostgreSQL instance (configured in the env file); now I'm setting up the db via Docker along with the schema migrations.

Since I already have the table schemas in JS, it shouldn't be too hard to extend this with some env variables so users can opt to use SQLite instead of Postgres given that this might not be user-friendly for full-time Windows users. That said, I'll need to do some research on how vectors/RAG/similarity search would work with SQLite, I haven't explored that yet.

In any case, I'm aiming to have a v0 prototype fully running by this weekend. If you're interested, I'd be happy to ping you once it's ready.

Also, just to clarify, I'm not affiliated with the project either, so I'll have to open a PR and hope the admin accepts it for this to be in production

[fly6516]

Thanks for the detailed follow-up! I'm sorry I was a bit busy yesterday, so although I saw your reply, I didn't have time to write my response.

I think having both Postgres (for full features) and SQLite (for lightweight / Windows-friendly setups) would cover most user scenarios nicely. For SQLite, I’ve seen some projects experimenting with sqlite-vec or similar extensions to enable vector similarity search—maybe that could be an option worth looking into. Moreover, most databases provide executable installation packages for Windows, making them relatively easy to handle. Therefore, there are still quite a few database options available.

Personally, I’d be very interested in trying out the v0 prototype once it’s ready, especially on Windows. I’d be happy to help test and share feedback on the setup experience.

By the way, the main reason I'm offering this suggestion is that some issues previously mentioned some problems with Linux rendering. If Docker support is also required, I believe it could make compatibility extremely complicated.

Looking forward to seeing how this evolves!

[lucas-oma]

Hey @fly6516 Sorry it took me some time to get back, you can either check the related pull request or directly on my fork's memory-system-feature branch

Any feedback is appreciated (Optimizations, if settings/config is not intuitive, improvements, etc)

[gg582]

#597

[gg582]

I tried to extend it from that PR, but I don't sure if it works

[fly6516]

Sorry for the late reply. Thank you for sharing the code. I've been busy with school and internship-related matters lately. I'll likely review and study it later. Much appreciated.

[konradre]

Should look at Anda-db and KIP protocol for agent memory