Thread No 1 of NeuraWeave
Perpetuity
Memory for AI agents.
A proxy that sits between your agent harness and its model. One base-URL change, and your agent runs through it.
Your agent's context never runs out. Each turn, the model gets a short frame of only the relevant context, about two thousand tokens, instead of the whole thread history. The frame grows when the relevant context needs more room. Nothing the agent has seen drops out of reach, and you stop paying for the whole conversation on every turn.
No model runs inside the memory loop. Nothing is extracted, summarized, or thrown away. A better retriever tomorrow reads more from the same memory.
Measured
- 92.4
- on LongMemEval-S, 500 questions, Opus 5 with native thinking, the production prompt, GPT-4o-2024-08-06, the official judge, nothing excluded
- 93.0
- with one 59M-token memory holding 500 users at once
- 99.1%
- of the evidence reaches the agent, measured on every question, no model on the path
- 84%
- fewer billed tokens across 200 real coding sessions, 5.6× in the median session, up to 16.4×
Every prompt, run and raw output is on the benchmarks page, with every caveat stated.
What it costs you
Every other memory system sends the thread history and adds its memory on top, so what a turn costs grows with the thread. Perpetuity replaces the thread history with the frame, so a turn costs about the same on day one and day ninety.
Across 200 real coding sessions that meant 84% fewer billed tokens, priced with prompt caching on both sides. $20 of model tokens does the work of up to $320, and of about $110 in a typical session. No session cost more.