Sonn — The memory that actually thinks

Sonn

Your AI agent has no memory.<br>So you've become its memory.

Close ✕

session-store · claude code<br>Sonn active

Sonn

Install Sonn →

How Sonn Works<br>Install Sonn →

Memory & reasoning for AI coding agents

The memory<br>that actually<br>thinks

Install Sonn →<br>built for claude code · more agents coming · local-first

Sonn remembers every decision, stops your agent repeating mistakes you've already fixed, and learns how you code. Built for Claude Code — and your code never leaves your machine.

Scroll

01 — The problem it solves

It catches the mistake<br>before your agent<br>makes it again.

A persistent process reasons over your history and steers before you ask — so settled decisions stay settled, instead of being relitigated every session.

→ Same prompt. Two outcomes.

session/cache.ts — claude code<br>no memory

1$claude "add caching to the session store"

3reaches for an in-memory TTL cache

4→ stale sessions across pods

5ships — you catch it in review. again.

✗ the bug you already fixed in May, written a second time.

session/cache.ts — claude code<br>Sonn active

1$claude "add caching to the session store"

3recall: we settled this in May← Sonn reasoned, before you asked

4use Redis with explicit invalidation

5the approach the team already agreed on

✓ caught before a single line was written. settled stays settled.

02 — Why it's different

Not another<br>memory tool.

Persistent memory is table stakes — everyone has it. The whole category just stores text and hands it back. Sonn is the layer above that: it reasons.

Other memory tools

store → embed → search → return

They compete on storage shape — vectors vs. graphs vs. paging. The memory itself never forms an opinion. They make the agent know more .

A filing cabinet.

Sonn

observe → reason → validate → steer first

A persistent process reasons over the memory — forms opinions about what matters and injects guidance before you ask . It makes the agent behave better .

A mentor.

03 — How it works

Three layers over<br>one local database.

One SQLite database on your machine, doubling as the vector store. The bottom layer is the commodity everyone ships. The two on top are the product.

03<br>WhisperBehavior

Your rules become trigger-gated nudges that fire the moment they're relevant — changing how the model behaves , not just what it knows. The closest thing to fine-tuning a model without its weights.

02<br>AngelReasoning

A persistent process that reasons over the memory — forms opinions, extracts patterns, and surfaces the right context before you ask . When it isn't confident, it says so and re-queries instead of returning noise.

01<br>MemoryStorage

Records every session, embeds it, retrieves it with hybrid search. The part every other tool stops at.

04 — Your code stays yours

Nothing of yours<br>runs on our servers.

Sonn is local-first. Your memory database lives on your machine, and generation runs through your own model credentials — your source goes to the provider you already trust, never to a third-party server . We hold the reasoning logic. You hold the code.

your code→ your machine

memory db→ your machine

generation→ your own credentials

reasoning logic→ Sonn (our IP)

0bytes of your source on our servers

05 — The numbers that matter

0.0%

LongMemEval — on a local model

426 / 470 · competitive with the best published results

lines of your code on our servers

local-first · your own trust boundary

Get started

Give your agent<br>a memory that thinks.

If you're reading this, you're early — and that means a lot to us. We're developers shipping the tool we always wanted. Try it, break it, tell us everything.

Install Sonn →<br>built for claude code · local-first · early access