Quartermaster Cuts Tool Calls 40% On The Same Tasks.

Quartermaster Case Study Architecture

Tool calls per task

QUARTERMASTER, CODEX
-40%
Without memory5
With Memorable3

Three replications, same result each time.

Pass rate

150 RUNS POOLED
91%
Without memory80%
With Memorable91%

Ahead in all three replications.

Quartermaster (QM) is an agent harness from YC that runs Codex in a Docker sandbox. Memorable stores procedures in QM's own Postgres. Every measured run used a fresh copy of the fixture, consent flipped read-only, and pass decided by sh test.sh run by the harness.

Five Tool Calls To Fix The Same Bug, Every Time.

Same three-bug fixture, fresh copy per run. Without a stored procedure Codex re-read the repo before every fix. In the five-family campaign, three of the four baseline failures quit after zero or one tool call.

The Stored Steps, Injected Every Turn.

The orchestrator injects the recalled procedure every turn, with a timeout and a size cap so it cannot block a run. The agent gets the steps that finished the task last time and skips the search.

Without memory

NO MEMORY
>Re-reads the repo

Five tool calls per task.

>Gives up early

Three of the four baseline failures quit after zero or one tool call.

>Pass 80%

Across every run.

With Memorable

PROCEDURAL MEMORY
>Follows the stored steps

Three tool calls per task.

>Keeps going

The pointer keeps the agent from quitting on the first miss.

>Pass 91%

Ahead in all three replications.

Adapter, Per-Turn Injection, Postgres.

{ 01 }

One adapter

One adapter over QM's core tool set captures every agent loop. Each tool call is joined to its result, so the procedure records what ran and what came back.

loop 1loop 2loop 3, 4one adaptercore tool setprocedurecall joined to result

Tool Calls Down 40%. Pass Rate 91%.

-40%

tool calls per task, 5 to 3

91%

pass with memory, 80% without

97%

pass across five task families with one shared store, against 87% without

EVAL_LOG.JSONEval log
MetricWithout memoryWith MemorableHow it was measured
Tool calls per task53Three replications
Pass rate80%91%Ahead in all three replications
Five task families, tool calls42One shared store. Recall picked the right family every time
Five task families, pass rate87%97%Order bugfix, config drift, crash-log triage, CSV repair, stale-caller rename

Try Memorable on your own agent

>npx memorable-cli@latest
># procedures stored in the Postgres you already run

Case Studies From Design Partners.