Local memory for AI agents

Every new chat
starts from zero.
Yours won't.

MemoryLayer is a memory graph that lives on your machine and every AI client reads and writes. Install it once — tomorrow's session already knows your stack, your decisions and your code.

Get started free

59% cheaper per question than letting an agent grep your repo — measured on real runs. See how we tested →

What happens today

Four things your agent does every session.

None of them are model problems. All four are the same missing piece.

01
Forgets you
Every chat starts blank
Preferences stay on this machine
02
Tools share nothing
Claude plans, Cursor codes
One namespace
03
Codebase invisible
Grep tax, every session
One call
04
Guesses with confidence
Wrong sounds like right
Checked before it speaks

How it works

One command. Then leave it on.

A SQLite graph on this machine. Setup wires every MCP client to it — nothing to export, nothing to re-paste tomorrow.

01
on this disk
One graph
SQLite on this machine. Not a cloud, not a paste buffer.
02
every client
Same namespace
Claude, Cursor, and Windsurf read and write the same file.
03
still on
No re-setup
Tomorrow's session starts from this graph. Leave it running.

this machine

~/.memorylayer

Local SQLite. Never uploaded.

npm i -g @raajvamsy/memorylayermemorylayer setup

Claude · Cursor · Windsurf

Measured

Same answers. A third of the bill.

One agent, one repo, 6 questions. The only thing we changed is whether it went hunting through the filesystem or asked memory once.

$0.0398$0.0162

average cost per question — 59% less

12,876 → 5,382 input tokens · shadcn/ui, 859 symbols

Read the full benchmark, including where memory loses →

On your disk

Nothing leaves the machine.

Runs locally
Memories, embeddings and code indexes live in SQLite on your own disk.
Zero content egress
The license server validates your key. It never receives a byte of memory content.
Checked before it speaks
verify() tests a claim against what's stored — grounded, unsupported, or refuted.

Pricing

Pay for calls, not seats.

Free

1,000 calls / month

Core memory + code tools
One-call code answers
Code indexing
Advanced memory ops
REST API (non-MCP clients)

Get started

Promost used

$19

5,000 calls / month

Core memory + code tools
One-call code answers
Code indexing
Advanced memory ops
REST API (non-MCP clients)

Get Pro

Pro+

$49

10,000 calls / month

Core memory + code tools
One-call code answers
Code indexing
Advanced memory ops
REST API (non-MCP clients)

Get Pro+

Every plan runs the same local graph. Compare every limit →

Questions

The ones people actually ask.

Does my data leave my machine?

No. Memories, embeddings and code indexes stay in local SQLite. The license server only validates your API key — it never receives memory content.

Which AI tools work with MemoryLayer?

Any MCP-compatible client — Claude Desktop, Cursor, Windsurf and others. Run memorylayer setup and it wires them up for you.

What's in the free tier?

1,000 tool calls a month, the core memory and code tools, and 50 recall(answer) calls. The full comparison has every limit.

How do I know an answer is actually correct?

Run verify() before your agent states it. Each claim comes back grounded, unsupported or refuted — a second opinion, not a promise.

Start

Stop re-explaining your project.

Set it up once, and Claude Code, Cursor and every other agent you use start each chat already knowing your decisions, fixes and code.

Get started free

install

npm i -g @raajvamsy/memorylayer

then memorylayer setup

Read the docs →

Local memory for AI agents

Every new chat
starts from zero.
Yours won't.

MemoryLayer is a memory graph that lives on your machine and every AI client reads and writes. Install it once — tomorrow's session already knows your stack, your decisions and your code.

Get started free

59% cheaper per question than letting an agent grep your repo — measured on real runs. See how we tested →

What happens today

Four things your agent does every session.

None of them are model problems. All four are the same missing piece.

01
Forgets you
Every chat starts blank
Preferences stay on this machine
02
Tools share nothing
Claude plans, Cursor codes
One namespace
03
Codebase invisible
Grep tax, every session
One call
04
Guesses with confidence
Wrong sounds like right
Checked before it speaks

How it works

One command. Then leave it on.

A SQLite graph on this machine. Setup wires every MCP client to it — nothing to export, nothing to re-paste tomorrow.

01
on this disk
One graph
SQLite on this machine. Not a cloud, not a paste buffer.
02
every client
Same namespace
Claude, Cursor, and Windsurf read and write the same file.
03
still on
No re-setup
Tomorrow's session starts from this graph. Leave it running.

this machine

~/.memorylayer

Local SQLite. Never uploaded.

npm i -g @raajvamsy/memorylayermemorylayer setup

Claude · Cursor · Windsurf

Measured

Same answers. A third of the bill.

One agent, one repo, 6 questions. The only thing we changed is whether it went hunting through the filesystem or asked memory once.

$0.0398$0.0162

average cost per question — 59% less

12,876 → 5,382 input tokens · shadcn/ui, 859 symbols

Read the full benchmark, including where memory loses →

On your disk

Nothing leaves the machine.

Runs locally
Memories, embeddings and code indexes live in SQLite on your own disk.
Zero content egress
The license server validates your key. It never receives a byte of memory content.
Checked before it speaks
verify() tests a claim against what's stored — grounded, unsupported, or refuted.

Pricing

Pay for calls, not seats.

Free

1,000 calls / month

Core memory + code tools
One-call code answers
Code indexing
Advanced memory ops
REST API (non-MCP clients)

Get started

Promost used

$19

5,000 calls / month

Core memory + code tools
One-call code answers
Code indexing
Advanced memory ops
REST API (non-MCP clients)

Get Pro

Pro+

$49

10,000 calls / month

Core memory + code tools
One-call code answers
Code indexing
Advanced memory ops
REST API (non-MCP clients)

Get Pro+

Every plan runs the same local graph. Compare every limit →

Questions

The ones people actually ask.

Does my data leave my machine?

No. Memories, embeddings and code indexes stay in local SQLite. The license server only validates your API key — it never receives memory content.

Which AI tools work with MemoryLayer?

Any MCP-compatible client — Claude Desktop, Cursor, Windsurf and others. Run memorylayer setup and it wires them up for you.

What's in the free tier?

1,000 tool calls a month, the core memory and code tools, and 50 recall(answer) calls. The full comparison has every limit.

How do I know an answer is actually correct?

Run verify() before your agent states it. Each claim comes back grounded, unsupported or refuted — a second opinion, not a promise.

Start

Stop re-explaining your project.

Set it up once, and Claude Code, Cursor and every other agent you use start each chat already knowing your decisions, fixes and code.

Get started free

install

npm i -g @raajvamsy/memorylayer

then memorylayer setup

Read the docs →

Four things your agent does every session.

Forgets you

Tools share nothing

Codebase invisible

Guesses with confidence

One command. Then leave it on.

One graph

Same namespace

No re-setup

Same answers. A third of the bill.

Nothing leaves the machine.

Pay for calls, not seats.

The ones people actually ask.

Stop re-explaining your project.

Four things your agent does every session.

Forgets you

Tools share nothing

Codebase invisible

Guesses with confidence

One command. Then leave it on.

One graph

Same namespace

No re-setup

Same answers. A third of the bill.

Nothing leaves the machine.

Pay for calls, not seats.

The ones people actually ask.

Stop re-explaining your project.