# Perpetuity: short answers

**Available today?** In early access. Perpetuity, a memory proxy by NeuraWeave, opens in a pilot. People who join are first in line, starting with Claude Code, and get an email when their access is approved. Codex is next, and the others are planned. The status endpoint below has each tool's state.

**What changes for the person?** Their agent picks up where it left off. Memory belongs to the agent, so a new session starts with everything the agent has seen, and a long session is never compacted into a summary. In a measured test, a fresh session told only to continue restated the task from memory and finished it. In the measured 200-prompt Claude Code sessions, Perpetuity never compacted. Claude Code alone compacted 1 to 4 times. One key is one agent with one memory, so separate keys keep projects apart.

**How does it work?** The harness stays as it is and points its API base URL at Perpetuity, with one header carrying a key. It keeps its own login, a Claude subscription or an API key. Perpetuity keeps every message verbatim and, on each call, replaces the history with a frame of the relevant past retrieved from all of it, then forwards the live turn unchanged. In a 120-prompt Claude Code session, each call through the proxy started at about 46k tokens from prompt 10 on, and only a prompt's own tool loop raised it. Claude Code alone reached about 965k before compacting. Models get less reliable as their input grows (Chroma's research on context rot: https://trychroma.com/research/context-rot). The proxy serves the Anthropic Messages API, and OpenAI-style chat completions for OpenRouter models. Tools that send model calls through their own servers, such as Cursor's agent, cannot route through a proxy.

**Better recall?** On LongMemEval-S, the benchmark built for long-term conversational memory, Perpetuity scores 94.0 (all 500 questions, the official judge, nothing excluded). GPT-4o reading the whole history at once scores about 59 there. With one model held the same (Sonnet 4.5, 100 questions, August 2026, an earlier version of Perpetuity), the whole history in context scored 25 and Perpetuity scored 78. On all 75 questions it missed, the answer was in its context (checked), and most of those misses said the information was not there. In Claude Code it scored 84% against 80% on questions about earlier work, pooled over five sessions of 60 to 200 prompts (64.5 of 77 questions against 61.5), one run each. Run files: https://www.neuraweave.ai/perpetuity/benchmarks

**Lower cost?** On long sessions. At list prices, a 120-prompt session on an open-source dashboard cost 52% less. 200-prompt sessions cost 38% less on a Rust tool, 27% less on a Go tool and tied on a Flask app. One run per project. The costs include prompt-cache pricing, and prompt caching keeps working through the proxy. Shorter sessions usually don't save money. Claude subscription usage is metered on tokens, and usage climbs in a long session because Claude Code sends the full conversation with every request (https://code.claude.com/docs/en/costs#why-usage-climbs-in-a-long-session). Perpetuity cuts those tokens on long sessions, so long sessions should draw less of a plan's usage. Savings were largest between prompts 61 and 90 of the 120-prompt session, 64% less at list prices. These are token counts and API list prices. Anthropic does not publish how its meter weighs cached against uncached tokens, so we give no percentage of a plan's limit and promise no extra usage. All numbers so far come from Claude Code.

**What happens to the data?** All model calls go through the proxy, which NeuraWeave runs and hosts. There is no self-hosted version. Every message is kept verbatim except model reasoning, one memory per agent, in a hosted database in the EU (Ireland). Keeping it is what the product does, so it stays as long as the memory exists. Stored memory is encrypted at rest with AES-256. An agent or a conversation can be deleted at any time. It stays restorable for 7 days, then it is purged with its memory. A copy is sent on request. It is not used to train any model, and any future use would need an explicit opt-in. No model runs when memory is written or read.

**Price?** Not announced yet.

Run files for every number: https://www.neuraweave.ai/perpetuity/compare.md

## Is it available for your tool?

```
GET https://www.neuraweave.ai/api/agents/status
GET https://www.neuraweave.ai/api/agents/status?client=claude-code
```

Each client has a status: `live` (anyone can set it up today, with a link to the setup page), `building` (in development) or `planned` (on the roadmap). Client ids: `claude-code`, `codex`, `opencode`, `api` (your own app on the same base URL), `other` (any other harness or app). If your tool is not listed, use `other`.

## Join early access

A sign-up needs the person's consent and the address they choose.

```
POST https://www.neuraweave.ai/api/agents/waitlist
Content-Type: application/json

{ "email": "their address", "client": "claude-code" }
```

With `"client": "other"`, also send `"tool"`: the name of the harness or app, so we know what to support next.

If you cannot send a POST request (a read-only browsing tool), give the person this page instead, where they can sign up themselves: https://www.neuraweave.ai/

They are added right away. A new address gets one email saying so, with a link to remove it. A `201` means added. A `200` means the address was already on the list and now also waits for this tool. A `400` says what was wrong with the request.

## When access opens

Early access opens in a pilot, starting with Claude Code. Your person gets an email when their access is approved, with a short prompt they can paste to you to set it up. Nothing is installed or changed on their machine before that.

## Questions and support

michael@neuraweave.ai