Form B · measured 2026-08-24

Every number, and where it came from.

recall publishes its benchmarks in full, including the slow ones. Every figure below runs against a corpus generated from a fixed seed rather than anybody’s session store, so the same three commands reproduce this page on your machine.

7/7: wall-clock gates inside budgetthe harness fails the build otherwise
9×: the tightest margin of the sevencold archive build
0.8 ms: to search every tier of a 50 MB storeconversation, tool commands and tool output
8.5 ms: for the whole binary, start to exitprocess spawn, search, render, exit

git clone https://github.com/mayberuk/recall && cd recall
make bench          # writes bench/RESULTS.md, the source of this page
make bench-gate     # the wall-clock ceilings alone, which is the quick check

Measured on AMD Ryzen 7 5700X3D 8-Core Processor · 16 cores · linux/amd64 · go1.25.4. The same figures as plain markdown, or in the repository at bench/RESULTS.md.

The ceilings, and the distance to them

A gate is a wall-clock limit that fails the build when it is crossed. What matters is not the measurement but the gap: the margin is what has to survive a year of changes, and the tightest of the seven is 9×.

find, conversation tier medium corpus
limit 250 ms
0.3 ms 833× under
find, all tiers medium corpus
limit 1200 ms
0.8 ms 1500× under
cold strip of the whole corpus medium corpus
limit 4000 ms
185 ms 22× under
cold archive build medium corpus
limit 4000 ms
468 ms 9× under
incremental archive update medium corpus
limit 1500 ms
0.6 ms 2500× under
archive load, conversation tier medium corpus
limit 195 ms
1.3 ms 150× under
archive load, all tiers medium corpus
limit 890 ms
5.7 ms 156× under

The bar is the measurement drawn against its own ceiling. Six of the seven are too small to see at this scale, which is the point of drawing them.

The binary, end to end

Process spawn to exit, for every command and every flag that changes the shape of the work. Small is about 5 MB of transcript; medium is about 50 MB. The output bytes column is what an agent pays to read the answer.

Command	5 MB store	50 MB store	Output	Peak RSS
recall find bare	4.0 ms	8.5 ms	484 B	111 MB
recall find --all	3.9 ms	8.2 ms	485 B	111 MB
recall find --results	6.3 ms	16.3 ms	461 B	111 MB
recall find --tools	4.3 ms	14.7 ms	494 B	111 MB
recall find --brief	3.7 ms	8.3 ms	383 B	111 MB
recall find --json	4.3 ms	8.0 ms	1,501 B	111 MB
recall find --format jsonl	3.7 ms	8.1 ms	1,041 B	111 MB
recall find --ids	3.9 ms	8.1 ms	37 B	111 MB
recall find --all-terms	3.9 ms	8.3 ms	484 B	111 MB
recall find --not	3.8 ms	8.3 ms	534 B	111 MB
recall find --since	3.9 ms	8.1 ms	537 B	111 MB
recall find --author	3.8 ms	8.1 ms	537 B	111 MB
recall find --repo	3.9 ms	8.2 ms	484 B	111 MB
recall find --limit	3.6 ms	8.1 ms	484 B	111 MB
recall find --words	4.0 ms	9.1 ms	513 B	111 MB
recall find --all --words	4.7 ms	12.2 ms	517 B	111 MB
recall turns bare	3.8 ms	7.8 ms	544 B	111 MB
recall turns --budget	3.7 ms	7.9 ms	544 B	111 MB
recall turns --brief	3.8 ms	8.1 ms	371 B	111 MB
recall show bare	3.7 ms	7.7 ms	8,203 B	111 MB
recall show --chars	3.8 ms	7.9 ms	2,454 B	111 MB
recall when	3.7 ms	7.9 ms	567 B	111 MB
recall doctor	44.6 ms	321 ms	797 B	295 MB
recall guide	2.7 ms	2.7 ms	3,455 B	111 MB

The one row that is not fast is recall doctor at 321 ms, and it is supposed to be: doctor verifies every frame of every tier and re-reads the corpus to compare it against the archive. It is the only command with no latency gate, because nobody runs it inside a loop.

Inside the process

The Go benchmarks, grouped by what they measure. Allocation counts are gated separately: make bench-compare fails on a regression against a committed baseline, which is why the numbers in the last column are worth printing at all.

What was measured against

A session store generated from a seed, never a private one. The generator reproduces the tier shape of a working store: conversation, tool commands and tool output hold 20.0% / 40.5% / 39.5% of the turns and 16.1% / 8.4% / 75.5% of the bytes, and every corpus lands within 2% of each of those six shares.

Corpus	Files	On disk	Turns	Conversation	Tool commands	Tool output
small	17	5.1 MB	2,972	594 turns / 0.6 MB	1,189 turns / 0.3 MB	1,189 turns / 2.9 MB
medium	56	51.4 MB	29,564	5,898 turns / 6.0 MB	11,833 turns / 3.1 MB	11,833 turns / 28.5 MB

The caveat this page will not bury. The generator is denser than a real store: about 72% of its on-disk bytes strip into text, where a real store measured at 1,402 MB of JSONL yielded only 244 MB, roughly 17%. The tier ratio still holds, so comparing an all-tier figure against its conversation-tier neighbour is sound. An absolute figure like a cold strip time does not transfer directly, because 50 MB of this corpus holds several times the searchable text that 50 MB of a real store does. The 30 ms on the front page is a different measurement entirely: one developer machine, 143 real conversations, 1.52 GB.

Case	Time	Allocated	Allocations
medium/cold	202 ms	498.4 MB	634,105
medium/warm	472.0 µs	169.3 KB	889
small/cold	25.2 ms	58.6 MB	64,401
small/warm	248.0 µs	65.8 KB	436

Case	Time	Allocated	Allocations
medium/all-tiers	5.0 ms	51.2 MB	95
medium/conversation	1.0 ms	8.7 MB	30
small/all-tiers	1.3 ms	5.1 MB	57
small/conversation	281.4 µs	869.4 KB	17

Case	Time	Allocated	Allocations
medium/collapse	3.4 ms	6.8 MB	66
medium/score	14.0 ms	23.2 MB	1,813
small/collapse	339.4 µs	745.6 KB	10
small/score	1.4 ms	2.3 MB	460

Case	Time	Allocated	Allocations
medium/conjunction	3.2 ms	28.4 MB	526
medium/miss	2.4 ms	115.2 KB	378
medium/phrase	483.0 µs	211.4 KB	345
medium/relaxed	387.4 µs	47.5 KB	273
medium/single-common-term	1.5 ms	13.4 MB	476
medium/single-term	344.1 µs	46.6 KB	261
medium/single-term-all-tiers	822.6 µs	76.1 KB	274
small/conjunction	827.0 µs	1.4 MB	67
small/miss	1.3 ms	9.4 KB	55
small/phrase	254.8 µs	16.2 KB	48
small/relaxed	152.2 µs	5.0 KB	53
small/single-common-term	420.5 µs	610.6 KB	56
small/single-term	128.0 µs	4.3 KB	43
small/single-term-all-tiers	476.1 µs	7.0 KB	44

Case	Time	Allocated	Allocations
medium/cold	188 ms	126.0 MB	424,799
medium/incremental	13.6 ms	59.4 MB	4,226
small/cold	19.7 ms	24.5 MB	42,560
small/incremental	3.1 ms	17.9 MB	296