I Built a Fully Browser-Native Vector Database (HNSW + IndexedDB, No Backend)

[Iconoclastdao]

Select Topic Area

Show & Tell

Body

Browser-Native Vector Database (HNSW + IndexedDB, No Backend)

Over the last two years I’ve been building a browser-native AI platform.
One of the core problems I ran into was semantic memory.

Most vector databases assume:

• A backend service
• Cloud infrastructure
• Persistent server processes
• DevOps overhead

That model doesn’t fit applications that are:

• Local-first
• Privacy-preserving
• Offline-capable
• Fully client-side

So I built a vector database that runs entirely in the browser.

No server. No external service. No hidden dependency.

---

What It Does

This system provides:

• HNSW indexing for approximate nearest neighbor search
• IndexedDB persistence for durability
• HOT / WARM / COLD tiered storage
• FP16 + int8 quantization to reduce memory footprint
• Metadata filtering
• AES-GCM encryption at rest
• Crash-safe journaling and rotation
• Lifecycle-aware auto-persistence
• Clustering support for semantic grouping

Search operates across memory tiers with fallback.
Data persists locally.
The system is designed to behave like infrastructure, not a demo.

---

Why This Approach

There are browser experiments and WASM ports of server systems, but many either:

• Keep everything in memory
• Rely on brute-force search
• Avoid durability concerns
• Ignore memory pressure
• Assume eventual server sync

I needed something that could act as a real semantic memory layer inside a browser-native platform.

The design priorities were:

• Zero infrastructure dependency
• Predictable memory behavior
• Local durability
• Privacy by default
• Practical performance within browser constraints

This is not a cloud replacement.
It’s an exploration of how far serious AI infrastructure can be pushed into the client.

---

Context

I transitioned into software about five years ago after working as a carpenter.
For the last two years I’ve been building a browser-native AI platform full-time.
This vector database is one of several internal components that emerged from that effort.

I’m sharing it because others building local-first AI systems may find it useful, or at least interesting from an architectural standpoint.

---

Repository

https://github.com/Semantic-Clustered-Memory-Protocol/SCMP

Feedback from people working on browser-native infrastructure or local AI systems is welcome.

[austinpatrew2]

This is seriously impressive running HNSW with IndexedDB fully client-side solves a real gap for privacy-first apps. I like how you rethought the architecture instead of defaulting to backend assumptions. It’s that same “cut the middle layer” mindset people look for when they want to sell house fast minnesota simplify the process and remove unnecessary overhead.

[Iconoclastdao]

I really appreciate the input. It means alot to me. I have been building for 2 years self taught last 5 years no validation if what i build is useful for others. Thank you...

…

On Sat, Feb 14, 2026 at 6:54 AM austinpatrew2 ***@***.***> wrote: This is seriously impressive running HNSW with IndexedDB fully client-side solves a real gap for privacy-first apps. I like how you rethought the architecture instead of defaulting to backend assumptions. It’s that same “cut the middle layer” mindset people look for when they want to sell house fast minnesota <https://cashforhousesmn.com/> simplify the process and remove unnecessary overhead. — Reply to this email directly, view it on GitHub <#187130 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BKR3BGELOASC3ZB6ORKQM6T4L4LITAVCNFSM6AAAAACVDY424SVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKOBQGQ4DAMA> . You are receiving this because you authored the thread.Message ID: ***@***.***>