rafinery

improve/improvements/resource-cache-unbounded.mdsee in the graph →

Frontmatter

schemaVersion: 1
id: resource-cache-unbounded
priority: P3
category: performance
status: open
title: _RESOURCE_CACHE grows unbounded for the life of the process
summary: Downloaded page content is memoized in a module-level dict that is never evicted or size-capped, so a long-running agent process accumulates every fetched URL's full markdown in memory
fix: Cap the cache (LRU / max entries) or add a size/TTL bound in download.py (~15 min)
leverage: { impact: low, effort: medium }
blast_radius: [agent-python, data-persistence]
cites:
  - agents/python/src/lib/download.py:17 :: _RESOURCE_CACHE
found: 2026-07-20
type: Improvement
description: "Downloaded page content is memoized in a module-level dict that is never evicted or size-capped, so a long-running agent process accumulates every fetched URL's full markdown in memory"
timestamp: 2026-07-20
tags: [performance, P3]

_RESOURCE_CACHE (download.py:17) memoizes each fetched URL's full HTML-to-markdown body keyed by URL, and is written on every successful (or errored) download (download.py:66,78,81). Nothing ever evicts it, so a process serving many research sessions grows in memory unbounded — the brain documents it as an in-process cache that "dies with the process," which is fine for a demo but is a slow leak under sustained use. Bound it (LRU with a max entry count, or a TTL) if this ever runs as a persistent service. Low priority for the example app; noted so the assumption is explicit.

Second reader since the 2026-07-27 refresh: fact_check_node now reads the same cache via get_resource rather than re-fetching, so the cache is load-bearing for correctness of the fact-check pass, not just a speed-up. Any eviction policy must not silently drop a resource the report still cites — evict by size/age, and treat a miss as "re-download", never as "unsupported claim".

Citations1verified at 1ebb712

agents/python/src/lib/download.py:17 · _RESOURCE_CACHE

Brain

Brain