memex-cli 0.2.1

# Token-size benchmark: memex context vs. git baselines

A comparison of context-payload sizes for three representative nodes. Approximate tokens at chars / 4 (close enough for an order-of-magnitude comparison; not exact tokenizer output).

This was the alternative to issue #38's full N=5 multi-arm benchmark. See `scripts/context-size.sh` for the script and the PR description for the methodology discussion.

## Method

For each node:

- **memex**: `cargo run -- context <id> --format markdown --depth 3`
- **git log (messages only)**: `git log --pretty=fuller -- <key_artifacts>` - what a non-memex user might scroll through.
- **git log -p (with diffs)**: same plus full diffs - the "everything available" upper bound an agent could be given.
- **PR body**: `gh pr view <num>` title + body for the PR matching the node's `git_ref`.

Reproduce with `scripts/context-size.sh <node-id>` from the repo root. Numbers captured 2026-05-03 against `main` at `44219c2`.

## Results

### Decision-heavy: `5a17fac4` - Audit and correct memex graph parent misassignments

Has 4 decisions and 3 rejected approaches in its memex summary.

| Source                              | Bytes  | Chars  | ~Tokens |
| ----------------------------------- | -----: | -----: | ------: |
| memex context --depth 3             |  3,759 |  3,735 |     933 |
| git log -p on key artifacts         | 34,640 | 34,603 |   8,650 |
| git log (messages only)             | 15,809 | 15,778 |   3,944 |
| gh pr #20 title + body              |  1,181 |  1,175 |     293 |
| **realistic baseline** (log + PR)   | 16,990 | 16,953 |   4,237 |

memex is **~22%** of the realistic baseline, and is the only source that contains the three rejected approaches (Reparent LICENSE to 269bfac9, leave the linear chain intact, and reparent d85b81 away from 970717, with each with its reason).

### Refactor with motivation: `b95077e3` - Remove graph.edges dual representation

Has 1 decision and 1 rejected approach.

| Source                              | Bytes  | Chars  | ~Tokens |
| ----------------------------------- | -----: | -----: | ------: |
| memex context --depth 3             |  2,630 |  2,626 |     656 |
| git log -p on key artifacts         | 81,605 | 81,461 |  20,365 |
| git log (messages only)             |  7,879 |  7,867 |   1,966 |
| gh pr #44 title + body              |    988 |    980 |     245 |
| **realistic baseline** (log + PR)   |  8,867 |  8,847 |   2,211 |

memex is **~30%** of the realistic baseline. The rejected approach ("add a check_integrity() call instead — Two representations are still maintained; integrity checks only catch divergence after the fact") is in the memex payload but absent from the commit message and the PR body.

### Shallow / negative control: `27ed3dfa` - Declare MSRV

Has 2 decisions and 0 rejected approaches. Memex shouldn't help much here.

| Source                              | Bytes  | Chars  | ~Tokens |
| ----------------------------------- | -----: | -----: | ------: |
| memex context --depth 3             |  2,862 |  2,854 |     713 |
| git log -p on key artifacts         |  3,984 |  3,980 |     995 |
| git log (messages only)             |  2,322 |  2,318 |     579 |
| gh pr #37 title + body              |    493 |    491 |     122 |
| **realistic baseline** (log + PR)   |  2,815 |  2,809 |     701 |

memex is **~102%** of the baseline, containing slightly *more* tokens with high content
overlap ("rust-version = 1.74 because clap 4 requires it" appears in both). For
shallow tasks the project-context preamble that memex always emits (root goal,
key decisions, key artifacts) is fixed overhead that dominates the payload.

## Caveats
- These three nodes don't provide a meaningful sample of possible use-cases. The point of this file is to make the comparison **reproducible and inspectable**, not to make a statistical claim. You can run the script on your own nodes and judge for yourself.
- Token counts are approximate (chars / 4). The realistic-baseline column adds log + PR; users may paste more or less than that. The ratios are what matter.
- All three nodes come from this repo, which is authored by a memex user. A non-memex repo would have different commit-message and PR-body length distributions, which affects the baseline.