Speak, don’t type
⌥Space — the shortcut Raycast and the Claude and Gemini desktop apps use — raises the window and toggles dictation. Speak, press it again, confirm.
whisper.cpp · large-v3-turbo · Metal · on-device
Voice-captured memory · DynamoDB vector search
EchoBrain
The Second Brain for Developers
Press ⌥Space and speak — or store data via agents like Claude CLI, Codex CLI, and Antigravity CLI. Three weeks later, ask for it in words you never used — and get it back.
No release is tagged yet — the button opens the releases page. Builds are ad-hoc signed, not notarised.
Say it
⌥Space opens the microphone. Speech is transcribed on your machine, by whisper.cpp over Metal — the audio never leaves it.
The problem
The context dies with the session
Gemini for Mac, the Claude and ChatGPT desktop apps, and terminal tools like Claude CLI, Codex CLI, and Antigravity CLI all share one limitation: you tell them something today and tomorrow they know nothing about it.
“We picked a DynamoDB vector index over OpenSearch because of the per-byte cost.”
Assistants forget
Every session starts empty. What you explained yesterday is gone, and you explain it again.
Notes need the right word
Full-text search only finds what you can already phrase. The note you cannot name is the note you cannot reach.
EchoBrain keeps it
What you say becomes a memory searchable by meaning — so “why didn’t we go with OpenSearch?” finds the sentence above, sharing almost none of its words.
In the app
Two seconds to capture. One question to get it back.
The transcript lands in the capture field for you to confirm before anything is stored — transcription misreads names, and a memory you never saw is one you cannot correct.
EchoBrain⌥Space
Features
Built for the way you actually think out loud
Everything here is in the shipped code. Nothing on this page is a mockup of a feature that does not exist yet.
Search by meaning
Recall is real vector search, not keyword matching. Every hit shows how close it actually was, so a bad result reads as a bad search rather than a confident wrong answer.
COSINE distance, inverted once, in the domain
Conversation mode
The microphone stays open. An endpointer decides when you stopped talking, and talking over the answer stops it mid-word — that one behaviour is what separates a conversation from a phone menu.
streaming frames · endpointing · barge-in
It knows whose voice
An enrolled voice print drops what the television said and stamps who spoke onto the memory. It is not a lock, so it is wired to exactly those two things — and pressing ⌥Space is never filtered.
off until you enrol it
Free unless you ask out loud
Capture and search cost one embedding. With no answer model configured — the default — the app never spends a token at all. Spoken questions are the one path that reaches a model.
a similarity floor gates the model, and sources are always shown
A hexagon the compiler enforces
The core crate has no Tauri, no cpal, no SQLite and no AWS. An accidental import fails the build instead of quietly eroding the boundary — and the frontend mirrors it, so the whole UI runs with no Rust toolchain and no AWS account.
core · audio · llm · app · ui
How it works
From a sentence to a neighbourhood
The app accesses and searches memories stored locally or recorded by terminal agents. For agents (Claude CLI, Codex CLI, Antigravity CLI) to access or store memories, installing the MCP server is required.
Microphone
cpal captures, rubato resamples to 16 kHz mono.
Transcription
whisper.cpp on Metal, entirely on your machine.
Routing
Cue phrases decide store or ask — before any model, so routing never costs a token. The guess shows as an editable chip.
Embedding
The sentence becomes a vector via the memory API.
DynamoDB
Vector and record land in the same item. Recall searches it.
Search is eventually consistent, so a capture appears immediately as “indexing” and a reconciler promotes it. A UI that hid the memory until the index caught up would just look broken.
Architecture
A client, not a backend
There is no backend code in this repository — no embedding code, no AWS SDK for storage. EchoBrain is a third driving adapter on the memory service port, next to the Lambda and the MCP server which integrates the same semantic memory into terminal agents (Claude CLI, Codex CLI, and Antigravity CLI) via the Model Context Protocol (MCP).
Lambda
HTTP API
MCP server
stdio, for agents (Claude CLI, Codex CLI, Antigravity CLI)
EchoBrain
macOS
MemoryService
the port, in agent-memory-core
HTTP + SigV4
API Gateway → Lambda → DynamoDB + Bedrock
The short dependency list in the core crate is the architecture. Components talk to a gateway interface, which is why the entire UI can be developed against fixtures with no network and no cost.
Install
Get it running
macOS today.
Download for macOS
macOS 13 Ventura or later · Apple Silicon or Intel
Read this before you download
No release is tagged yet, so the button opens the releases page. When one lands, the build is ad-hoc signed and not notarised — Gatekeeper will refuse it on any Mac but the one that built it. You can bypass this by removing the macOS quarantine attribute with `xattr -d com.apple.quarantine EchoBrain.app` (and granting executable permission with `chmod +x`), though building from source remains the supported path until notarisation is set up.
Configuration
A file rather than environment variables: an app opened from the Finder inherits no shell profile. Every field is optional — leave the answer model out and the app never spends a token.
config.toml
endpoint = "https://memory-mcp-example.execute-api.us-east-1.amazonaws.com/"
region = "us-east-1" # the API's region — this is what SigV4 signs for
profile = "rfsales"
github_login = "dmux" # attribution only, never authorizationStatus
Where we are
Verified end to end against the deployed API. Clippy clean with warnings denied, 174 tests, static export building.
0Walking skeletondone
Workspace, core, UI and Tauri wired to the real API.
1⌥Space dictationdone
cpal → rubato → whisper-rs on Metal.
1.5Voice answersdone
Spoken answers over Bedrock, record-or-ask routing, settings.
2.5Conversation modedone
Streaming frames, endpointing, barge-in, follow-up window.
2Journal and librarydone
SQLite journal, library view, reconciler, rating and deactivation.
3Native panel
NSPanel, vibrancy, menu-bar icon, animation.
4Wake word
And the embedding model behind speaker identity.