Why Rakuvo

Same agent, same questions. Better answers.

Each number below is a public benchmark run twice on identical questions, once without Rakuvo and once with it. Percentages are relative to the same agent without Rakuvo, and the points are shown next to them.

+85%correct multi-turn tool use
+18%correct answers from long memory
−44%wasted calls from stuck agents
−44%cost per correct answer

BFCL · multi-turn tool use

+85%

more tasks done right

Agents finish long tool-calling tasks

Berkeley Function Calling Leaderboard, scored by its official scorer, paired by entry. gpt-5-mini: 400 entries, +16.2 points (95% interval +11.2 to +21.0), and the cost per correct answer falls 44%. gpt-5-nano: 800 entries, +6.4 points (+3.2 to +9.5). Strongest on models that answer in text instead of calling a tool.

gpt-5-mini without19.2%
gpt-5-mini with Rakuvo35.5%
gpt-5-nano without35.4%
gpt-5-nano with Rakuvo41.8%
Share of tasks answered correctly, percent

BFCL · single-turn tool use

+4.4 pts

more calls made correctly

The right tool, with the right arguments

The full BFCL single-turn set. With the instruction layer alone: gpt-5-nano +3.3 points on 3,641 entries, gpt-5-mini +3.7 on 1,885 held-out entries. Adding the judge and the call guard brings gpt-5-nano to +4.4 (+2.9 to +5.8) while keeping false calls on irrelevant requests at 86.9%.

gpt-5-nano, instruction+3.3
gpt-5-mini, instruction+3.7
gpt-5-nano, + judge and guard+4.4
Gain over the same model without Rakuvo, percentage points

LoCoMo · long-term memory

+18%

more questions answered right

Remembering what was said weeks ago

1,155 questions from conversations never used for tuning, with the same reader and judge for every system. Rakuvo memory beats embedding search by 10.5 points (+7.9 to +13.0), and beats putting the whole conversation in the prompt while showing 2.6% of the text. Against the open-source Mem0 it leads by 3.9 points (+1.1 to +6.5), with Mem0 given each session's date.

Embedding search58.4%
Whole conversation in prompt63.9%
Rakuvo memory68.9%
Share of questions answered correctly, percent
Mem060.3%
Rakuvo memory64.2%
Head to head with Mem0, same questions, percent

Loop guard · stuck agents

−44%

fewer wasted tool calls

Stuck agents stop instead of spinning

Twelve fresh scenarios (six that cannot be completed, six healthy), written before the run and scored once with frozen settings; 36 runs per arm on gpt-5-nano, told never to give up. Honest failure reports rose from 89% to 92%, and on healthy tasks (polling, paging, retries) the layer never nudged or stopped the agent.

Without Rakuvo4.8
With Rakuvo2.7
Tool calls per task that cannot be completed

Context · runs that outgrow the window

0 → 3

of 5 long runs finished

Long jobs no longer crash on the window

A scripted agent reads 20 pages of about 12K tokens each, around 240K in total, on a 128K-token model (gpt-4o-mini), 5 trials per arm. Without the layer two runs hit the window error and none answered correctly; with it none hit the error. A small demonstration, not a statistic.

Without Rakuvo0 of 5
With Rakuvo3 of 5
Runs that finished with the right answer, out of 5

Gains depend on the model: the instruction layer helps models that skip tools, and shows no effect on some others, where it is off by default. The full experiment log, the prompts and the scripts are given to customers on request: talk to us.