You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 912dae6
Browse filesBrowse the repository at this point in the historyBrowse files
Copy file name to clipboardExpand all lines: README.md
+2-2Lines changed: 2 additions & 2 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -470,9 +470,9 @@ Published backend benchmarks on Apple M3 Pro. These are separate from the CBM Le
470
470
471
471
**RAM-first pipeline**: All indexing runs in memory (LZ4 HC compressed read, in-memory SQLite, single dump at end). Memory is released back to the OS after indexing completes.
472
472
473
-
**Token efficiency, within the reported workloads**: The project's [original README report](https://github.com/DeusData/codebase-memory-mcp/blob/4ff6feeab639c4668da0f491930326ed5ec63093/README.md#performance) lists ~3,400 tokens for five structural queries via codebase-memory-mcp versus ~412,000 via file-by-file exploration: approximately **120× fewer tokens** for that comparison. The separate [31-repository preprint](https://arxiv.org/html/2603.27277v1#S4.SS1) reports **10× lower token use** and **2.1× fewer tool calls**, with answer-quality scores of **83% for CBM versus 92% for file-by-file exploration**. These are scoped backend measurements, not a promised saving for every task or a benchmark of CBM Lens. Actual savings depend on the question, repository, agent, and required verification. The study's [system-performance measurements](https://arxiv.org/html/2603.27277v1#S4.SS3) include the Linux indexing and Cypher timings above.
473
+
**More focused context, fewer tokens**: The project's [31-repository preprint](https://arxiv.org/html/2603.27277v1#S4.SS1) reports **10× lower token use** and **2.1× fewer tool calls** compared with file-by-file exploration. These March 2026 backend measurements cover the study's workloads and agent setup; the paper documents the full evaluation. An earlier [five-query comparison](https://github.com/DeusData/codebase-memory-mcp/blob/4ff6feeab639c4668da0f491930326ed5ec63093/README.md#performance) reported ~3,400 versus ~412,000 tokens for structural queries. The study's [system-performance measurements](https://arxiv.org/html/2603.27277v1#S4.SS3) include the Linux indexing and Cypher timings above.
474
474
475
-
To measure comparable quality, latency, and agent-efficiency metrics on your own workload, see [Measuring quality, latency, and agent savings](docs/MEASURING_SAVINGS.md). Exact reproduction of the figures above requires the original inputs and raw artifacts.
475
+
Explore the [measurement guide](docs/MEASURING_SAVINGS.md) to evaluate quality, latency and token use on your own workload.
<ahref="images/lens-architecture.png" aria-label="View the full CBM Lens architecture screenshot"><imgsrc="images/lens-architecture.png" width="1440" height="900"alt="CBM Lens architecture map showing the source folders of codebase-memory-mcp" fetchpriority="high"></a>
347
+
<ahref="images/lens-overview.gif" aria-label="View the full CBM Lens architecture walkthrough"><picture><sourcemedia="(prefers-reduced-motion: reduce)" srcset="images/lens-architecture.png" width="1440" height="900"><imgsrc="images/lens-overview.gif" width="960" height="600" alt="CBM Lens architecture walkthrough: explore the source folders, open src/pipeline, select artifact.c and read its source" fetchpriority="high"></picture></a>
348
348
<figcaption>CBM Lens, exploring this repository. Real product captures from October 2026. <ahref="https://github.com/DeusData/codebase-memory-mcp/blob/main/docs/images/README.md">Capture details ↗</a></figcaption>
349
349
</figure>
350
350
<divclass="facts" aria-label="Project at a glance">
<figure><ahref="images/lens-galaxy.jpg" aria-label="View the full Lens Galaxy screenshot"><imgsrc="images/lens-galaxy.jpg" width="1280" height="720"alt="CBM Lens Galaxy graph for exploring symbols and relationships" loading="lazy"></a><figcaption><strong>Follow the connections</strong>Search and expand the Galaxy graph.</figcaption></figure>
402
+
<figure><ahref="images/lens-galaxy.gif" aria-label="View the full CBM Lens Galaxy walkthrough"><picture><sourcemedia="(prefers-reduced-motion: reduce)" srcset="images/lens-galaxy.jpg" width="1280" height="720"><imgsrc="images/lens-galaxy.gif" width="960" height="600" alt="CBM Lens Galaxy walkthrough: search a function, expand its outgoing calls and follow a path through the graph" loading="lazy"></picture></a><figcaption><strong>Follow the connections</strong>Search, expand connections and follow paths.</figcaption></figure>
403
403
<figure><ahref="images/lens-system-structure.png" aria-label="View the full Lens system structure screenshot"><imgsrc="images/lens-system-structure.png" width="1440" height="900" alt="CBM Lens system structure showing source groups and their dependencies" loading="lazy"></a><figcaption><strong>Inspect system structure</strong>See source groups and dependencies.</figcaption></figure>
404
404
<figure><ahref="images/lens-chat.png" aria-label="View the full Lens local chat screenshot"><imgsrc="images/lens-chat.png" width="1440" height="1200" alt="CBM Lens optional local chat explaining selected source using indexed facts" loading="lazy"></a><figcaption><strong>Ask about selected code</strong>Optional browser AI, with on-device inference.</figcaption></figure>
<divclass="evidence" id="benchmark"><div><pclass="eyebrow">Performance and research</p><h2>More useful context.<br>Less repeated work.</h2></div><div><pid="token-savings">Reuse the graph to find relevant symbols, follow relationships and retrieve focused source snippets. Compact, paged results help keep repeated repository exploration out of your agent’s context window.</p><p>Our <ahref="https://arxiv.org/abs/2603.27277v1">March 2026 preprint</a> reports <strong>10× lower token use</strong> and <strong>2.1× fewer tool calls</strong> across 31 repositories compared with file-by-file exploration. Reported answer-quality scores were 83% for Codebase Memory and 92% for the file-exploration agent.</p><p>These historical backend results predate CBM Lens. Savings vary with the task, repository, agent and source verification needed.</p><divclass="evidence-links"><ahref="https://github.com/DeusData/codebase-memory-mcp#performance">Results and limitations ↗</a><ahref="https://github.com/DeusData/codebase-memory-mcp/blob/main/docs/BENCHMARK.md">Benchmark methodology ↗</a><ahref="https://arxiv.org/abs/2603.27277">Research paper ↗</a></div></div></div>
467
+
<divclass="evidence" id="benchmark"><div><pclass="eyebrow">Performance and research</p><h2>More useful context.<br>Less repeated work.</h2></div><div><pid="token-savings">Reuse the graph to find relevant symbols, follow relationships and retrieve focused source snippets. Compact, paged results help keep repeated repository exploration out of your agent’s context window.</p><p>Our <ahref="https://arxiv.org/abs/2603.27277v1">March 2026 preprint</a> reports <strong>10× lower token use</strong> and <strong>2.1× fewer tool calls</strong> across 31 repositories compared with file-by-file exploration.</p><p>These measurements cover the study’s backend workloads from March 2026. See the paper for the evaluation setup and full results.</p><divclass="evidence-links"><ahref="https://github.com/DeusData/codebase-memory-mcp#performance">Token-use results ↗</a><ahref="https://github.com/DeusData/codebase-memory-mcp/blob/main/docs/BENCHMARK.md">Indexing benchmarks ↗</a><ahref="https://arxiv.org/abs/2603.27277">Research paper ↗</a></div></div></div>
468
468
</section>
469
469
470
470
<sectionclass="section" id="faq">
471
471
<divclass="faq"><pclass="eyebrow">Common questions</p><h2>Is Codebase Memory<br>right for your workflow?</h2>
472
472
<details><summary>When should I use Codebase Memory?</summary><p>Choose CBM when you want your coding agent and your own source-navigation workspace to use the same local graph of code, docs and infrastructure. It brings structural search, source evidence, CBM Lens, agent integrations and experimental test selection into one open-source package. Native releases run on macOS, Linux and Windows without Docker or a language runtime. <ahref="#features">Explore the capabilities</a> and <ahref="#smart-ci">Smart CI’s current scope</a>.</p></details>
473
-
<details><summary>How does Codebase Memory reduce AI token use?</summary><p>Your agent can query indexed relationships and retrieve the relevant symbols or source snippets instead of repeatedly exploring whole files. Compact results and pagination control how much context each query returns. The benefit depends on the task; <ahref="#token-savings">read the measured token savings and quality tradeoff</a>.</p></details>
473
+
<details><summary>How does Codebase Memory reduce AI token use?</summary><p>Your agent can query indexed relationships and retrieve the relevant symbols or source snippets instead of repeatedly exploring whole files. Compact results and pagination control how much context each query returns. The benefit depends on the task; <ahref="#token-savings">explore the published token-use measurements</a>.</p></details>
474
474
<details><summary>Does Codebase Memory send my source code anywhere?</summary><p>The backend indexes source and stores and queries its graph locally. Your coding agent has its own data-handling settings. Lens can optionally download a browser model after you enable it, with inference on-device. Installers and update checks use the network to obtain releases and assets.</p></details>
475
475
<details><summary>How does the index stay current?</summary><p>Incremental indexing updates changed files. With auto-watch enabled, an eligible active MCP session can watch an already indexed project. Automatic first indexing is a separate opt-in setting. The local UI daemon is not a promise of a permanent background watcher.</p></details>
476
476
<details><summary>How much language coverage is available?</summary><p>Bundled tree-sitter grammars cover a broad range of languages and formats. Supported language families also have Hybrid LSP semantic resolution. Extraction depth, type resolution and framework relationships vary; the <ahref="https://github.com/DeusData/codebase-memory-mcp#language-support">language support matrix</a> describes the details.</p></details>
<aclass="source-link" href="https://github.com/DeusData/codebase-memory-mcp/releases">View all releases on GitHub →</a>
488
488
</section>
489
489
</main>
490
-
<footer><divclass="container footer-inner"><div><aclass="brand" href="#main">Codebase Memory</a><p>Local code intelligence. Open source, MIT licensed.</p></div><divclass="footer-links"><ahref="https://discord.gg/KguGXxg2X"><imgsrc="https://img.shields.io/badge/Discord-Join_the_community-5865F2?logo=discord&logoColor=white" alt="Join the community on Discord" width="187" height="20" loading="lazy"></a><ahref="https://github.com/DeusData/codebase-memory-mcp">GitHub</a><ahref="https://github.com/DeusData/codebase-memory-mcp#readme">Documentation</a><ahref="https://github.com/DeusData/codebase-memory-mcp/blob/main/SECURITY.md">Security</a><ahref="llms.txt">llms.txt</a></div></div></footer>
490
+
<footer><divclass="container footer-inner"><div><aclass="brand" href="#main">Codebase Memory</a><p>Local code intelligence. Open source, MIT licensed.</p></div><divclass="footer-links"><ahref="https://discord.gg/KguGXxg2X"><imgsrc="https://img.shields.io/badge/Discord-Join_the_community-5865F2?logo=discord&logoColor=white" alt="Join the community on Discord" width="187" height="20" loading="lazy"></a><ahref="https://github.com/DeusData/codebase-memory-mcp"><imgsrc="https://img.shields.io/github/stars/DeusData/codebase-memory-mcp?style=social&label=Star%20on%20GitHub" alt="Star Codebase Memory on GitHub" height="20" loading="lazy"></a><ahref="https://github.com/DeusData/codebase-memory-mcp#readme">Documentation</a><ahref="https://github.com/DeusData/codebase-memory-mcp/blob/main/SECURITY.md">Security</a><ahref="llms.txt">llms.txt</a></div></div></footer>
Copy file name to clipboardExpand all lines: docs/llms.txt
+2-2Lines changed: 2 additions & 2 deletions
Original file line number
Diff line number
Diff line change
@@ -25,7 +25,7 @@
25
25
26
26
## Measurement and limits
27
27
28
-
A March 2026 preprint across 31 real-world repositories reports 10× lower token use and 2.1× fewer tool calls compared with file-by-file exploration, with answer-quality scores of 83% for CBM and 92% for file-by-file exploration. These historical results describe the study's workloads and agent setup; they are not a universal percentage saving or a promise of lower cost for every session. The README Performance section preserves the comparisons and their limits.
28
+
A March 2026 preprint across 31 real-world repositories reports 10× lower token use and 2.1× fewer tool calls compared with file-by-file exploration. These historical backend measurements cover the study’s workloads and agent setup. The paper documents the full evaluation; the README Performance section links to the research and earlier measurements.
29
29
30
30
Graph evidence depends on index generation, freshness, and exact path/scope coverage. Partial, skipped, excluded, stale, or unknown coverage requires source verification, especially for absence claims. Infrastructure relationships are static, and test-selection support is narrower than language-indexing support.
31
31
@@ -43,7 +43,7 @@ Graph evidence depends on index generation, freshness, and exact path/scope cove
0 commit comments