# Stop paying to resend the whole conversation, agent history or coding session

Stop paying to resend the same conversation. Every turn resends the whole thread; Tileward Context sends only the part that answers. This page covers how much less it sends on real usage, what happens to answers and speed, and how to install it.

Page: https://tileward.com/context/

## Headline results

- [95% fewer tokens](https://tileward.com/research/a-nineteenth-of-the-context/) on long conversations, with most answers kept.
- [68% fewer tokens](https://tileward.com/research/a-third-of-the-context/) on agent work, with the same work done.
- [40–59% fewer made-up answers](https://tileward.com/research/a-nineteenth-of-the-context/#no-answer) when the conversation can't answer.
- [19× sooner](https://tileward.com/context/#speed) for the model to start responding, and 18.6× more requests per GPU at a 3 s wait for it to start.

## Measured on our own usage

Read off the same dashboard a customer gets:

- 90.7% of one long-running Claude Desktop thread was never resent: 21.6M tokens of 23.8M.
- 79.7% median cut per conversation on Claude Code sessions over 50k tokens (45–97% across those 61); counting all 179 sessions, short ones included, the median is 63%.
- 89.1% across every conversation, lifetime: 23.3M saved of 26.1M.

If you already run prompt caching, read 90.7% against a cached baseline: a cache hit is a discount on tokens you still send, which still occupy the window, while Tileward's are not sent at all. Against a well-cached baseline the dollar saving shrinks by roughly an order of magnitude while the tokens stay withheld.

## Quality

Every turn stays stored as it was said; for each question, Tileward sends the turns that bear on it, in the words that were used. The console shows which turns were sent for each question and what they cost.

- In our own test, a model given only those turns answered every question it could answer from the whole conversation, from a fraction of the tokens ([the test](https://tileward.com/evals/#long-conversation)).
- When the answer isn't in the conversation, it makes one up 26% of the time, against 63% with the whole transcript ([how we measured it](https://tileward.com/research/a-nineteenth-of-the-context/#no-answer)).

## Speed

A model reads everything it was sent before it writes a word, and that takes longer as the history grows. A 255k-token history sent as 51k takes about 13× less GPU time to read, 49 s against 3.9 s.

| History sent to Tileward 35B-A3B | Wait until it starts responding |
|---|---|
| 8k tokens | 0.42 s |
| 32k tokens | 2.1 s |
| 64k tokens | 5.4 s |
| 130k tokens | 15.6 s |
| 203k tokens | 33 s |
| 255k tokens | 49 s |

On a model that answers without reasoning first, the whole reply lands about 20% sooner, even against a cached prompt.

## Estimate your own

The page's calculator applies the share measured on each kind of flow to your volume: agents in sequence, 49% fewer (2.0×); one agent with a long backlog, 68% fewer (3.2×); parallel gatherers, 78% fewer (4.5×). It counts input tokens only, at the list rates on the [pricing page](https://tileward.com/pricing/); the more of your input a prompt cache serves, the smaller the saving.

## Install: three steps

1. [Create a free account](https://app.tileward.com/?join=explore): no card, with $25 of credit, 2,000 recalls and 100 MB of documents.
2. Pick your client in the [connect guide](https://tileward.com/docs/#connect). Context is an MCP server, so it plugs into the client you already use: a few lines to paste, signed in through the browser or with a key.
3. Ask it to remember something, then recall it a turn later. From then on, your thread's savings read off the console dashboard.

The [docs](https://tileward.com/docs/) carry the endpoint table, which authentication each client can do, and the conversation id to send with every call. Tileward Governance runs against the same key. Context is included in every Tileward account; you pay for model usage at each model's rate ([see the rates](https://tileward.com/pricing/)).

## Related

- [Long conversations](https://tileward.com/research/a-nineteenth-of-the-context/): the write-up behind the long-conversation and made-up-answer figures
- [Agent work](https://tileward.com/research/a-third-of-the-context/): the write-up behind the agent figure
- [Evals](https://tileward.com/evals/): every measured result, one line each
- [Talk to an engineer](https://calendar.app.google/hg2DzFTxKJgwFWm19): bring a long conversation and see what it retrieves