From f3a0e624576ae8934dd3b2cb4472a290df749621 Mon Sep 17 00:00:00 2001 From: saddestmartian Date: Wed, 1 Jul 2026 14:31:01 -0700 Subject: [PATCH] docs: add real-world bash/ops-tooling repo example to Performance The existing Performance section benchmarks large polyglot application codebases (Linux kernel, Django). Bash already scores in the Excellent parsing tier under Language Support, but there's no worked example of the agentic token-efficiency payoff on a repo shape that's mostly shell scripts, YAML/JSON config, and Markdown docs rather than a typical multi-language application. Adds one real session's measurement: a single search_code call vs a grep-based sub-agent fan-out on an internal dev-tooling repo, clearly labeled as one real-world data point rather than a controlled benchmark. Signed-off-by: saddestmartian Signed-off-by: saddestmartian --- README.md | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/README.md b/README.md index c1c17f9f8..e97704a7e 100644 --- a/README.md +++ b/README.md @@ -231,6 +231,14 @@ Benchmarked on Apple M3 Pro: **Token efficiency**: Five structural queries consumed ~3,400 tokens via codebase-memory-mcp versus ~412,000 tokens via file-by-file grep exploration — a **99.2% reduction**. +### Real-World Example: A Bash/Config-Heavy Ops-Tooling Repo + +The benchmarks above are measured on large polyglot application codebases (Linux kernel, Django). It's worth showing what "Excellent"-tier Bash parsing (see [Language Support](#language-support)) translates to for a different repo shape: infra/ops-tooling repos that are mostly shell scripts, YAML/JSON config, and Markdown docs — not a typical multi-language application. + +One real session on an internal dev-tooling repo (bash + JSON + Markdown, ~14.9k indexed nodes / ~20.5k edges): a coding agent's single `search_code` call reproduced — ranked and deduplicated — the full list of consumers of a shared JSON config file in **~375ms**. The same task, done by fanning out a grep-based sub-agent instead, took **~131s across 12 tool calls** to assemble the equivalent answer by hand. Shell functions indexed as first-class graph nodes with `CALLS` edges made this possible on plain tree-sitter extraction alone — no Hybrid LSP tier needed. + +*(One real-world measurement from an actual agent session, not a controlled multi-trial benchmark — offered as a data point for teams wondering whether the token-efficiency value proposition holds outside typical application codebases.)* + ## Troubleshooting & Diagnostics codebase-memory-mcp runs **100% locally and collects no telemetry** — your code, queries, environment, and usage never leave your machine. That privacy guarantee also means that when you hit something we can't reproduce on our side (a slow memory climb over hours, a performance regression, a leak that only appears after days of real use), **we have no data at all unless you choose to send it.** Here is how to capture it yourself.