Acervo Benchmark Report

2026-03-27 09:23 — v0.2.2-3 — 6 scenarios, 360 turns

76%

Average savings

360

Total turns tested

94%

Avg context hits

Phantom entities

165

Entities extracted

Token usage over time — Saas Founder Real (101 turns) (101 turns)

Without Acervo (red), every message adds to the context. With Acervo (green), the knowledge graph keeps token usage flat regardless of conversation length.

91%

Savings at turn 101

~486 tk

Acervo context

5,243 tk

Baseline at turn 101

Scorecard — all 6 scenarios

Token savings and context hit rate side by side. Green = savings, blue = context hit rate (% of turns where Acervo injected relevant context).

Token anatomy — what Acervo sends vs what it would cost

Sampled every ~10 turns from Saas Founder Real (101 turns). The stacked bars show Acervo's composition (warm context from graph, hot layer, system+user). The red bars show what full history would cost.

Per-scenario breakdown

Scenario	Turns	Savings	Context Hits	Entity Recall	Nodes	Edges	Phantoms

Estimated API cost per scenario

Based on published API pricing (per 1M tokens). Output estimated at 30% of input tokens.

Notable moments

Auto-detected events: context restored after long gaps, cost milestones, peak compression, graph growth.

Scenario	Turn	Type	What happened

Token trends per scenario

Acervo (green) vs full history (red) for each scenario individually.

Conversation evidence

Actual messages, injected context, and extracted entities for each turn.