• Sources: primary
  • Summary: Loom stores agent memories, a repository call graph index, embeddings, full-text index and run logs in one ClickHouse database, and flags a memory for review when the code it describes is rewritten. On ClickHouse code questions the mean graded score rose from 0.78 to 0.84 and cost per question fell from $0.212 to $0.17, and file-finding accuracy rose from 91.4% to 96.4%. It ships under Apache 2.0 with Claude Code and Codex plugins, MCP, HTTP and SDKs, and needs ClickHouse 26.5 or later.
  • Why it matters: Teams running several coding agents get a self-hosted, namespace-scoped memory that Claude Code and Codex sessions can share, with a published benchmark method.
  • Background: Coding agents start each session with an empty context, so they rediscover repository quirks or guess. Memory tools store notes from past sessions but usually cannot see the code, so notes go stale silently. Loom combines vector, BM25 and entity-graph searches and keeps every version of a fact so queries can run as of a past date.