Rubi: memory for AI coding agents
Collection
Knowledge graph and context manager for Claude Code, Codex, Gemini CLI, Cursor and local models. Interactive calculator plus the 90-session benchmark. • 2 items • Updated
repo stringclasses 2
values | braco stringclasses 3
values | rep int64 1 3 | pergunta int64 1 3 | custo_usd float64 0.09 0.23 | turnos int64 2 14 | tempo_s int64 14 50 | input int64 4 18 | cache_escrita int64 17.6k 30.1k | cache_leitura int64 61.2k 384k | output int64 599 3.77k | tokens_equiv int64 44.5k 113k | erro float64 | nos_no_grafo_ao_fim float64 0 6 ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
flask | A | 1 | 1 | 0.155403 | 6 | 38 | 10 | 25,482 | 193,826 | 1,469 | 77,702 | null | null |
flask | A | 1 | 2 | 0.102107 | 3 | 21 | 6 | 18,359 | 109,293 | 680 | 51,053 | null | null |
flask | A | 1 | 3 | 0.125553 | 6 | 25 | 10 | 18,281 | 191,945 | 1,402 | 62,777 | null | null |
flask | C | 1 | 1 | 0.155147 | 4 | 34 | 8 | 28,348 | 155,095 | 1,072 | 77,574 | null | 0 |
flask | C | 1 | 2 | 0.133778 | 7 | 28 | 8 | 21,450 | 158,310 | 1,630 | 66,889 | null | 0 |
flask | C | 1 | 3 | 0.17082 | 6 | 40 | 12 | 24,200 | 253,031 | 2,339 | 85,410 | null | 0 |
flask | D | 1 | 1 | 0.147261 | 5 | 30 | 6 | 28,138 | 110,585 | 1,258 | 73,631 | null | 5 |
flask | D | 1 | 2 | 0.136566 | 7 | 27 | 8 | 22,056 | 157,479 | 1,683 | 68,283 | null | 5 |
flask | D | 1 | 3 | 0.127758 | 3 | 29 | 6 | 21,806 | 111,260 | 1,827 | 63,879 | null | 5 |
flask | A | 2 | 1 | 0.154328 | 6 | 32 | 8 | 26,841 | 143,692 | 1,821 | 77,164 | null | null |
flask | A | 2 | 2 | 0.114132 | 4 | 22 | 8 | 18,526 | 152,712 | 947 | 57,066 | null | null |
flask | A | 2 | 3 | 0.192522 | 9 | 50 | 14 | 24,030 | 293,620 | 3,765 | 96,261 | null | null |
flask | C | 2 | 1 | 0.124797 | 3 | 21 | 4 | 26,316 | 61,876 | 715 | 62,399 | null | 0 |
flask | C | 2 | 2 | 0.09683 | 2 | 18 | 4 | 19,464 | 62,882 | 639 | 48,415 | null | 0 |
flask | C | 2 | 3 | 0.138927 | 3 | 24 | 4 | 27,546 | 62,876 | 1,616 | 69,464 | null | 0 |
flask | D | 2 | 1 | 0.130865 | 3 | 24 | 4 | 27,184 | 63,206 | 948 | 65,433 | null | 5 |
flask | D | 2 | 2 | 0.11187 | 3 | 19 | 6 | 20,522 | 111,652 | 744 | 55,935 | null | 5 |
flask | D | 2 | 3 | 0.110347 | 3 | 18 | 4 | 21,637 | 64,203 | 1,095 | 55,173 | null | 5 |
flask | A | 3 | 1 | 0.225032 | 14 | 50 | 18 | 30,089 | 383,650 | 2,791 | 112,516 | null | null |
flask | A | 3 | 2 | 0.088967 | 2 | 14 | 4 | 17,606 | 61,173 | 630 | 44,483 | null | null |
flask | A | 3 | 3 | 0.173786 | 7 | 44 | 10 | 25,815 | 200,731 | 3,036 | 86,893 | null | null |
flask | C | 3 | 1 | 0.153596 | 5 | 32 | 8 | 27,635 | 155,398 | 1,196 | 76,798 | null | 0 |
flask | C | 3 | 2 | 0.175807 | 9 | 34 | 14 | 23,332 | 297,055 | 2,304 | 87,904 | null | 0 |
flask | C | 3 | 3 | 0.117379 | 3 | 22 | 6 | 20,575 | 112,337 | 1,260 | 58,690 | null | 0 |
flask | D | 3 | 1 | 0.144676 | 5 | 27 | 6 | 27,662 | 110,580 | 1,190 | 72,338 | null | 5 |
flask | D | 3 | 2 | 0.112425 | 3 | 17 | 6 | 20,496 | 114,645 | 750 | 56,213 | null | 5 |
flask | D | 3 | 3 | 0.127632 | 3 | 31 | 6 | 22,307 | 111,209 | 1,615 | 63,816 | null | 5 |
hono | A | 1 | 1 | 0.142154 | 7 | 26 | 8 | 25,270 | 144,241 | 1,221 | 71,077 | null | null |
hono | A | 1 | 2 | 0.108331 | 5 | 19 | 6 | 19,276 | 104,526 | 1,031 | 54,166 | null | null |
hono | A | 1 | 3 | 0.111203 | 5 | 21 | 6 | 19,698 | 104,596 | 1,148 | 55,602 | null | null |
hono | C | 1 | 1 | 0.133282 | 4 | 22 | 6 | 25,959 | 107,368 | 796 | 66,641 | null | 0 |
hono | C | 1 | 2 | 0.112974 | 4 | 17 | 6 | 20,810 | 108,811 | 796 | 56,487 | null | 0 |
hono | C | 1 | 3 | 0.114943 | 5 | 20 | 6 | 20,924 | 108,626 | 951 | 57,472 | null | 0 |
hono | D | 1 | 1 | 0.141051 | 4 | 24 | 6 | 27,652 | 110,806 | 827 | 70,526 | null | 6 |
hono | D | 1 | 2 | 0.107553 | 3 | 15 | 4 | 21,664 | 68,994 | 709 | 53,776 | null | 6 |
hono | D | 1 | 3 | 0.106424 | 2 | 20 | 4 | 20,987 | 64,638 | 954 | 53,212 | null | 6 |
hono | A | 2 | 1 | 0.126322 | 5 | 25 | 6 | 24,421 | 99,078 | 881 | 63,161 | null | null |
hono | A | 2 | 2 | 0.111342 | 4 | 20 | 8 | 18,545 | 147,479 | 765 | 55,671 | null | null |
hono | A | 2 | 3 | 0.119887 | 6 | 19 | 8 | 19,460 | 153,056 | 1,142 | 59,944 | null | null |
hono | C | 2 | 1 | 0.135013 | 4 | 26 | 6 | 26,512 | 107,614 | 743 | 67,506 | null | 0 |
hono | C | 2 | 2 | 0.113002 | 4 | 19 | 6 | 20,701 | 113,080 | 757 | 56,501 | null | 0 |
hono | C | 2 | 3 | 0.113111 | 4 | 19 | 6 | 20,498 | 112,985 | 851 | 56,556 | null | 0 |
hono | D | 2 | 1 | 0.127678 | 3 | 21 | 4 | 27,085 | 63,649 | 660 | 63,839 | null | 6 |
hono | D | 2 | 2 | 0.120848 | 4 | 20 | 6 | 22,179 | 112,201 | 968 | 60,424 | null | 6 |
hono | D | 2 | 3 | 0.109522 | 3 | 20 | 4 | 21,054 | 64,638 | 1,237 | 54,761 | null | 6 |
hono | A | 3 | 1 | 0.12753 | 5 | 20 | 6 | 24,438 | 103,480 | 907 | 63,765 | null | null |
hono | A | 3 | 2 | 0.08975 | 3 | 17 | 4 | 17,555 | 65,562 | 641 | 44,875 | null | null |
hono | A | 3 | 3 | 0.111777 | 5 | 20 | 6 | 19,614 | 104,844 | 1,234 | 55,888 | null | null |
hono | C | 3 | 1 | 0.132154 | 4 | 18 | 6 | 26,144 | 107,882 | 599 | 66,077 | null | 0 |
hono | C | 3 | 2 | 0.13153 | 6 | 25 | 8 | 21,976 | 159,698 | 1,167 | 65,765 | null | 0 |
hono | C | 3 | 3 | 0.100732 | 3 | 19 | 4 | 19,569 | 67,641 | 892 | 50,366 | null | 0 |
hono | D | 3 | 1 | 0.141081 | 4 | 20 | 6 | 27,702 | 110,856 | 809 | 70,541 | null | 6 |
hono | D | 3 | 2 | 0.120969 | 3 | 19 | 6 | 22,490 | 116,334 | 773 | 60,484 | null | 6 |
hono | D | 3 | 3 | 0.106727 | 3 | 17 | 4 | 21,366 | 69,026 | 745 | 53,364 | null | 6 |
Rubi benchmarks
Does a project memory make Claude Code cheaper, and which savings are real once their own cost is counted? This dataset holds every run of the benchmarks behind Rubi: 174 Claude Code sessions on two public repositories (Flask and Hono), plus the pilots. Model claude-sonnet-5, fresh clone per run, answers checked automatically wherever the task has a ground truth.
Rounds 1 and 2: reading and writing memory (90 sessions)
Three chained questions per repository, each in a fresh session.
| Arm | What runs |
|---|---|
| A | Claude Code, no plugin |
| B | Rubi with harvest after every answer that explored 4 or more files (cost of writing memory) |
| C | Rubi with map and automatic recall, empty graph (cost of reading) |
| D | Same as C, starting with a graph of 5 or 6 nodes written by Claude (mature project) |
| Per repetition, Flask and Hono summed (6 sessions) | A | C | D |
|---|---|---|---|
| Turns | 34.0 | 26.7 (−21%) | 21.3 (−37%) |
| Cache reads | 919k | 771k (−16%) | 579k (−37%) |
| Wall-clock time | 161 s | 146 s | 133 s (−17%) |
| Billed cost | $0.793 | $0.785 (−1%) | $0.744 (−6%) |
Writing memory after every answer (arm B) cost +15% in these short sessions.
Round 3: automatic graph built without any model (36 sessions)
Same questions as rounds 1 and 2, clones with 400 commits of history. Arm C has an empty memory with the automatic graph off; arm E adds what Rubi builds by computation alone: file top comments in the text index, automatic module summaries, and edges between files that change together in git.
| E vs C, paired (n = 6) | Mean change | 95% CI |
|---|---|---|
| Turns | −19.0% | −36.9% to −1.1% |
| Billed cost | −9.6% | −19.3% to +0.1% |
Answer quality by key files cited: 18/18 (C) and 17/18 (E).
Long sessions: output shortening × trimmed context (24 sessions)
One session doing four tasks in a row on Hono, 2×2 design, 6 repetitions per cell. All 96 answers correct.
| Factor (12 pairs) | Billed cost | 95% CI |
|---|---|---|
| Shortening large Bash outputs | −9.9% | −22.8% to +3.0% |
| Denying 4 unused tools (−4.9k tokens per call) | +13.7% | −6.6% to +34.0% |
Short sessions: output shortening (24 sessions)
Four tasks with large outputs, 3 repetitions: −1.8% (95% CI −9.4% to +5.8%), 12/12 correct in both arms.
Files
rodada-1-escrita.csv,rodada-2-leitura.csv,rodada-3-automatico.csv,economia-1.csv,saidas-1.csv: one row per session (cost, turns, tokens by type, time, correctness where checked)estatistica-economia-1-saidas-1.md: paired statistics*-piloto.*: pilot runs, kept for transparency*.md: per-round summariesrespostas/: every answer, per arm, repetition and questionsementes/: the seed graph used by arm Dtarefas.json: the questions of rounds 1 to 3
Reproduce in the Rubi repository: node bench/rodar.mjs, node bench/economia.mjs, node bench/saidas.mjs, then node bench/estatistica.mjs economia-1 saidas-1. Full discussion in the README.
- Downloads last month
- 140