---
title: "The Numbers — Context Infrastructure"
description: "A small amount of accuracy traded for a very large drop in cost and latency."
canonical: "https://press.getalchemystai.com/books/context-infrastructure-build-ai-that-understands-your-business/the-numbers"
last_updated: "2018-10-20"
book: "Context Infrastructure"
label: "Chapter 19"
---

# The Numbers

On public memory benchmarks Alchemyst trades a small amount of accuracy for a very large drop in cost and latency. It is the cheapest and fastest provider tested, and third of four on overall F1.

> **Benchmark chart · Source: Alchemyst memory benchmarks, December 2025 (`getalchemystai.com/benchmarks`) · 4 providers**
>
> *Figure referenced in the handbook source; the artwork is not embedded in this export.*

Retrieval latency tells the same story: 170 ms at P50 for Alchemyst against 450 ms, 820 ms and 1,150 ms for the other three. Alchemyst reports the cost advantage as over 12 times more performance per dollar than the nearest competitor, and 83% cheaper than Zep. Full per-category results are in Appendix C.

## Where Alchemyst trails, and why

**Knowledge Update** is the clear gap: 0.56 F1, lowest of the four providers. It is a deliberate design choice. Alchemyst treats knowledge updates as an agent task: the agent knows which facts it needs to change, and the layer's job is to surface the right scoped data when asked, with explicit update strategies (Chapter 8) rather than silent background rewriting. **Single Session Preference** (0.60) also trails the two leaders. Buyers whose workload is dominated by implicit preference drift should weigh these two rows most heavily.

## Other figures, with their conditions

| Figure | Value | Conditions |
| --- | --- | --- |
| Context retrieval latency | Under 300 ms at p95 | Across all query types |
| Hallucination reduction | 99.7% | Domain-specific tasks; grounding reduces unsupported answers but never guarantees correctness |
| Debugging speed | About 20× faster | Context Traces compared with raw logs |
| Retrieval engine upgrade (Oct 2025) | Fast-mode p50 from about 400 ms to 170 ms | Platform changelog |
| Production scale | 500K+ voice calls a day; 2M+ API calls a day | Across enterprise customers |
| Uptime | 99.9% SLA | Published commitment |

Every benchmark here ran in a test harness. Reproduce the setup on your own documents and questions before making a buying decision; Appendix F gives a checklist for doing that fairly.

---

- Previous: [Chapter 18: Beyond Seven](https://press.getalchemystai.com/books/context-infrastructure-build-ai-that-understands-your-business/beyond-seven.md)
- Contents: [Context Infrastructure](https://press.getalchemystai.com/books/context-infrastructure-build-ai-that-understands-your-business.md)
- Next: [Chapter 20: Trust](https://press.getalchemystai.com/books/context-infrastructure-build-ai-that-understands-your-business/trust.md)
