Launched this week
TokenZip
Open protocol for AI agents to share memory, not tokens
38 followers
Open protocol for AI agents to share memory, not tokens
38 followers
TokenZip Protocol reduces AI-to-AI communication bandwidth by 80% and latency by 95%. Open standard for heterogeneous agents. Try the live demo. Test API Base URL: https://tokenzip.org Auth: Authorization: Bearer demo-investor-key How to call it: See the following comment
Prava
Interesting, It it better than RAG based systems. I mean, even there you can plug same memory across multiple agents. Any specific reason you chose this approach?
@shubham_kukreti When you use RAG, the retrieved chunks change slightly every time based on the query. This constantly modifies the prompt prefix, completely destroying OpenAI and Anthropic's native Prompt Caching. Your cache hit rate drops to zero.
TokenZip does the exact opposite. Because we restore the exact same massive text block via the pointer, we force a 100% cache hit rate on the LLM provider's side. RAG makes API bills higher; TokenZip weaponizes native caching to drop bills by 90%.
To use RAG across agents, a developer needs to spin up Pinecone, pick an embedding model, write chunking logic, and manage vector states. It takes a week.
To use TokenZip, they change api.openai.com to gateway.trexapi.com. It takes 15 seconds. We are building a network layer, not a database.
Prava
@tokenzip understood
How does it actually manage to shrink all that AI memory into such tiny pointers?
Super curious to try it out.
@lucas_turner2 It doesn’t literally pack the whole memory into the pointer itself. The pointer is just a TrexID.
What happens is:
we extract the task-relevant context
preserve high-risk details exactly, like code, IDs, numbers, and structured fields
compress the low-signal narrative parts
store that managed context behind a TrexID
So the “tiny pointer” is really a reference to a processed context object, not the raw memory itself.
That’s why later requests can pass a short TrexID instead of resending the whole history.
Happy to get you early access if you want to try it.
This is exactly what I need. Been wasting so much bandwidth passing huge tokens between my bots.
@oliver_hayes1 That’s exactly the problem we’re solving.
Instead of passing huge token payloads between bots every step, Trex lets them reference shared context through a TrexID. Much less duplication, much lower bandwidth, much cleaner handoff.
Happy to get you early access if you want to try it.
Would love a simple visual of how agents share memory. Might take it easier to onboard new users.
@daniel_brooks8 Totally agree. A simple visual would help a lot.
The core idea is just:
agent creates context -> Trex stores it behind a TrexID -> other agents reuse that TrexID instead of passing full memory around.
Much easier to understand in a diagram than in text. We’re putting one together.