01
Learn
Capture situation → reply pairs from your real AI conversations.
Every exchange with your coding agent becomes a structured memory of context and your authentic response.
Open Source • MIT License • Local First
The shadow of you.
Shadow Link learns how you react to AI coding agents, retrieves relevant past responses, and predicts what you are likely to say next.
// process
How it works
A closed loop between your behavior and the agent — observed, weighted, and reflected back.
02
Match
Find related reactions using lexical memory, optional embeddings, and time-decayed weights.
Relevance is scored across language, semantics, and recency so the freshest, closest moments surface first.
03
Predict
Surface an editable ghost reply shaped by how you actually communicate.
A draft in your own voice appears inline — accept it, edit it, or reject it. Every choice trains the mirror.
// capabilities
Built like an instrument.
Terminal-native components engineered for memory, reflection, and reinforcement.
Local memory graph
A connected store of situations, reactions, and outcomes — queried entirely on your machine.
Claude Code live-session mirror
Observes your active coding session in real time and reflects your patterns back as they happen.
Editable ghost replies
Drafts in your voice you can accept, rewrite, or dismiss inline.
Feedback & reinforcement loop
Every accept or edit reweights the corpus toward how you really respond.
Temporal memory decay
Recent reactions carry more weight; stale habits gracefully fade.
MCP server integration
Exposes memory and prediction as Model Context Protocol tools any compatible agent can call.
Optional Claude Code proxy
Route sessions through a local proxy to capture richer context — bound to localhost by design.
Local LLM & embedding support
Plug in on-device models for inference and embeddings — nothing leaves your hardware.
// the interface
Your workflow.
Your workflow.
Your memory.
Your voice.
A terminal-native TUI that lives where you already work — no browser, no tab, no telemetry leaving the box.
- Animated command palette in pure green-on-black
- Live telemetry: corpus size, latency, decay window
- Session picker across every captured conversation
- Ghost-confidence meter before you accept a reply
match --context "rename handler?"
↳ retrieved 3 reactions · decay window 14d · score 0.91
▸ ghost reply · editable
"yeah rename it to handleRequest and keep the old export as an alias so nothing breaks downstream."
confidence91%
# press ⏎ accept · e edit · x reject — every choice trains the mirror
corpus 4,812 pairs
latency 38ms
embeddings local
decay 14d
mcp :7420
▸ install from source
$ git clone https://github.com/ThatReid23/Shadow-Link-v3.git
$ cd Shadow-Link-v3
$ python -m venv .venv
$ .\.venv\Scripts\Activate.ps1
$ pip install -e .
$ shadow-link daemon
$ shadow-link ui
// get started
Install in
Install in
under a minute.
Clone the repository, create a virtual environment, install the package, then launch the daemon and the terminal UI. No accounts, no cloud setup.
View full documentation →
// trust model
Local by default.
Your reaction corpus stays in a local SQLite database. Embeddings and chat inference remain local unless you explicitly configure remote endpoints.
Proxy mode uses mitmproxy to observe sessions and should remain bound to
localhost. Nothing is exposed to the network unless you deliberately change that.