# Perpetuity

[94.0 on LongMemEval-S, 500 questions. See the benchmarks](https://www.neuraweave.ai/perpetuity/benchmarks)

keeps your coding agent from summarizing away your session
while using fewer tokens

With Perpetuity

45k tokens read per call only what the prompt needs, about the same size every call

Without Perpetuity

739k tokens read per call the whole history, re-sent every call

- Claude Code Soon
- Codex
- + more
- API Soon

[Follow the thread](https://www.neuraweave.ai#story)

Agents have

## Amnesia

Sessions end. Compaction deletes. Summarized memory is lossy.

Agent context is

## Expensive

The whole conversation history is re-sent every turn. Most of it is not needed for the current prompt.

Reasoning

## Degrades

with context

The same agent scores lower with the full conversation history than with a retrieved frame.

Every message

## Kept

For 99.1% of questions the evidence reaches the agent within one recall call, for 97.4% already in the first frame. Measured on the 470 questions that have evidence.

Your budget goes

## Up to 2.8× further

52% less spent on tokens in a 120-prompt Claude Code test, the same agent side by side. 64% less between prompts 61 and 90.

Reasoning

## 94.0

from a 3.6k-token frame

LongMemEval. Opus 5 with thinking, the official GPT-4o judge, 500 questions, nothing excluded. Half of them answered in one round.

## 92.2 at 82M-token context

Holding 500 users at once

### Context that never runs out

Each turn, the model gets a frame of only the relevant context, about ten thousand tokens in a coding session, instead of the whole conversation history. The agent can search its memory when it needs more.

### Nothing thrown away

No model runs inside the memory loop, and nothing is extracted or summarized. A better retriever tomorrow can read more from the same memory.

Measured in Claude Code sessions, the same agent side by side

52%

less spent on tokens over a 120-prompt coding session, $50.24 against $104.71

84%

of questions about earlier work answered correctly, 64.5 of 77 over 5 sessions, against 80% for Claude Code on its own after compacting its history

45k

tokens read per call by prompt 60, most of it Claude Code's own system prompt and tools. Without Perpetuity, 739k.

[How it works](https://www.neuraweave.ai/perpetuity/how-it-works)

## Perpetuity opens in early access.

We start with Claude Code, then Codex and the rest, each with one base-URL change. Join now and you are first in line. We email you when your access is approved.

This page as HTML: https://www.neuraweave.ai

Perpetuity for agents (availability, and joining early access with the person's consent): https://www.neuraweave.ai/perpetuity/agents.md