the morning re-explain
Every session starts blank. An hour of re-pasting context before any real work begins.
r
relay
pre-seed · 2026
Your thinking,
permanently yours.
A persistent memory layer that runs across every AI tool you use. Decisions, constraints, and context, portable, owned by you, never lost on a tool switch.
// the problem02 / 10
Your AI tools don't talk
to each other. You're the bridge.
Conversations are stateful within a session and stateless between them. Switch tools, open a new chat, the state dies. It isn't a bug in one product. It's how every AI tool on the market is built today.
decisions go missing
Last week's rejected approach, last month's constraint, gone with every new chat window.
silent drift
Friday's agent never heard Tuesday's constraint. It ships the wrong thing. No warning.
// the cost03 / 10
You are the
context router.
You generate handoff document after handoff document. You paste them at the start of new sessions. You watch them degrade when they exceed context limits. You lose architectural decisions that took hours to reach.
It delays you. It introduces errors. It's a job no one should have to do.
$ open new-tool
> context: EMPTY
$ paste handoff-v7.md
> 14,200 tokens …
> truncated, exceeds limit
$ re-explain constraints
> 3 decisions lost
λ …again tomorrow▌
MANUAL OPS~30 MIN / DAYEVERY TOOL
// the solution04 / 10
A vendor-neutral layer
that sits over every AI tool.
relay is a persistent layer owned by you, written to and read from by any AI tool. Your reasoning lives in one place, separate from the tools you think with. Not another walled garden. A layer below all of them.
Captures reasoning
Decisions, constraints, and rejected approaches, stored with full provenance, not raw transcripts.
Loads anywhere
Drop your full context into Claude, Cursor, ChatGPT, or Gemini in under a second.
Stays yours
If a tool raises prices or ships a worse model, you switch without starting over. It was never theirs.
⊙ how it works05 / 10
Two commands. Save, then relay.
01
relay save
At the end of any session, capture every decision, constraint, and rejected approach, with the reasoning behind each. No manual notes.
02
relay load [project]
At session start, drop your full context into any tool, decisions, live constraints, indexed rejections. The tool is ready before your first prompt.
λ relay load auth
> Constraint: no JWT, drivers run offline
> Decision: device-local token store
> Rejection indexed: JWT, session #7
✓ Context loaded · 12 artifacts active
λ▌
MEMORY LAYER<300MSANY TOOL
// v1 architecture & scope06 / 10
Honest about the mechanism.
V1 ships what works today, keyword search, manual invocation, native MCP for compatible tools. The "it just knows" vision is the destination, not the launch claim. We show V1 and ship V1.
v0.2.1 · shipping
Local MCP + SQLite
Local MCP server over stdio. 3 typed tools, SQLite + FTS5 search. relay setup wires Claude Desktop.
v0.3.0 · next
Chrome extension loop
Localhost HTTP + Chrome extension. Captures ChatGPT, Gemini, Claude.ai right from the browser.
v0.5+ · vision
Remote MCP + cloud
Remote MCP → ChatGPT + cross-device (v0.5). Cloud + Pro (v0.6). Intelligence layer (v0.7).
scope is the right size: a real tool for the user who feels the pain now, not a demo of a future that doesn't exist yet.
// why now07 / 10
The window is real.
And it's narrow.
The pain is hitting mainstream consciousness, Yale SOM is publishing "Who Owns Your AI Memory?"; Indie Hackers and Hacker News threads are active. The number of AI tools per workflow is climbing every quarter, and every new tool widens the gap.
Memory tools today are developer APIs (Mem0, Cognee), not user-facing products. That gap is real now, but well-funded players have the integrations to move into it. We have 12–18 months to reach the users they can't learn from.
4+AI tools per power-user workflow
$24Mraised by the nearest dev-API player
12–18mowindow before they go user-facing
0user-facing, vendor-neutral products today
// the wedge08 / 10
Built for the
person who is
already the bridge.
The wedge is the multi-tool power user, engineers, researchers, and builders who run four AI tools a day and feel context loss as a daily tax. The strongest signal: the founder is that user, doing relay's job by hand right now.
tools in play · daily
Claude · Anthropic
Cursor · IDE
ChatGPT · OpenAI
Gemini · Google
// validation plan09 / 10
Phase 0: prove the pain.
Before building, we validate. A stealth, researcher-curious landing page positioned as independent research, not a product pitch. It asks one question: "I see this pattern in my own AI workflow. Do you?"
capture
Email signups
From people who confirm they feel the pain, the first ranked list of real demand.
validate
Survey
Frequency, tools used, and the biggest pain in their own words, qualified, not guessed.
recruit
1:1 interviews
Opt-ins for user research that shape what V1 actually ships.
distribution · founder-led on Twitter/X · Reddit · Ask HN · 1:1 outreach, every channel researcher-curious, never a pitch