Your agents forget. You shouldn’t re-explain.
Design in Claude. Build in Codex. Review in Grok. Radeon is the brain that sits outside every model — so context follows the work, not the tab.
# finish a session in Claude Coderadeon save "auth before payments · use Postgres"✓ 3 decisions saved to payments-api# next morning, you open Codexradeon load payments-api→ packed 1.2k tokens · decisions, prefs, open questions# Codex already knows. no re-brief.▍
Same prompt. Same second agent. One of them already knows.
add Stripe webhooks — match how we already do auth
which auth pattern? let me look…
$ grep -rn "auth" . → 214 matches
sessions or JWT? where do payments live? what did we decide last time?
You become the human cache — re-typing architecture, preferences, and decisions you already made.
add Stripe webhooks — match how we already do auth
Using your JWT + Postgres pattern from payments-api. Auth middleware runs before handlers, per the decision you made last session.
The agent opens already knowing the project. No re-brief. You pick up mid-thought.
Every agent starts from zero
You already pay for the smartest models on earth. Switching tools still means re-briefing architecture, preferences, and open questions — every single time.
Context dies at the tab
What you decided in Claude never reaches Codex. You become the human cache, re-typing the same architecture between tools.
Chat history isn't knowledge
Agents need decisions and constraints — not a wall of transcripts dumped back into the prompt and hoping the model re-derives intent.
Vendor memory doesn't travel
Built-in product memory keeps you inside one lab. Continuity across tools is nobody's default job — so it never happens.
Similarity isn't understanding
Vector top-k finds nearby text. It won't resolve conflicts, respect what changed, or hand off task state cleanly.
Capture. Structure. Hand off.
Radeon sits between your agents. It remembers what matters for the project — and packs it for whichever model you open next.
Capture the work
Session extracts, explicit pins, and handoff notes from Claude, Codex, Cursor, and Grok. High-signal only — never every keystroke.
Structure the brain
Decisions, preferences, open questions, procedures — typed, scoped to one project, with sources and time attached.
Pack & continue
Open another agent. Radeon injects a tight, budget-aware context pack for the task. Same brain. Zero re-brief.
Built for multi-agent builders
Not a chatbot. Not a company-wide knowledge dump. A portable project brain that lives outside every model.
Typed memory
Facts, prefs, decisions, procedures, open questions — each with scope and provenance. Not blobs.
Session handoff
End on one agent, continue on another with summary, decisions, and next steps intact.
Context packs
Budget-aware packing for the task at hand. Signal in the window — not everything you ever said.
Inspectable brain
See exactly what Radeon knows. Edit it. Delete it. Trust requires visibility.
Model-agnostic
MCP and native hooks for Claude, Codex, Cursor, Grok, and more. Your brain isn't locked to a lab.
Forgets on purpose
Supersede, resolve conflicts, decay. A wrong memory is worse than no memory at all.
Outside every model. Inside every session.
Agents call Radeon. Radeon stores project memory as a small linked graph — then packs the right neighborhood for the next session.
Memory is a graph: decisions, prefs, and open questions linked to the project. Packs pull the neighborhood you need — not the whole history.
Get early access
Private alpha for builders who switch agents daily and are done re-explaining the same project.
No spam. One email when your access is ready.