w1n5t0n/OptMem
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

the wake budget is 208 lines: 16k tokens measured, not 16k assumed

Browse source 
WAKE_LINES=256 was converted from Taelin's 16k-token budget at 4 bytes per
token. Dense technical memories run 3.36, so the real document was 19,015
tokens on the live store -- 19% over the number he set, and it needed a
fourth tool call to deliver a 17-line stub.

  256 lines  63,922 bytes  19,015 tokens  4 parts
  208 lines  54,011 bytes  16,393 tokens  3 parts   <- shipped

Nothing is recomputed: the budget only selects which existing lines print.
test.py now reads the budget from memo instead of keeping its own copy.
This commit is contained in:
victortaelin 2026-07-25 18:49:37 -03:00
parent 909e28fec9
commit a6abc264c5
3 changed files with 9 additions and 9 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

12
README.md
View file

@ -96,11 +96,11 @@ keeps a block whole when its size is small relative to its age, so **detail is
proportional to recency**, and it spends exactly `WAKE_LINES` lines doing it: proportional to recency**, and it spends exactly `WAKE_LINES` lines doing it:
``` ```
10,000 memories, WAKE_LINES = 256: 10,000 memories, WAKE_LINES = 208:
block size: 1 2 4 8 16 32 64 128 256 block size: 1 2 4 8 16 32 64 128 256
how many: 54 27 27 27 28 27 27 27 12 how many: 42 21 21 21 22 21 21 21 18
└ the last 54, verbatim ───────────▶ the first 3,000, 256:1 └ the last 42, verbatim ───────────▶ the first 4,600, 256:1
``` ```
The oldest memories are recalled as a vague shape, the newest word for word, The oldest memories are recalled as a vague shape, the newest word for word,
@ -138,7 +138,7 @@ decisions; drop wording.
``` ```
bad worked on the memory system today and made good progress on the design bad worked on the memory system today and made good progress on the design
good OptMem design settled: LOG.txt append-only truth, TREE binary merge good OptMem design settled: LOG.txt append-only truth, TREE binary merge
tree of 280-char summaries, wake renders a fixed 256-line document tree of 280-char summaries, wake renders a fixed 208-line document
``` ```
## Output is delivered in parts ## Output is delivered in parts
@ -152,7 +152,7 @@ piece:
Codex 10,000 tokens (configurable per call) Codex 10,000 tokens (configurable per call)
``` ```
A 256-line memory is ~64 KB, so a single-shot `memo wake` is mangled A 208-line memory is ~56 KB, so a single-shot `memo wake` is mangled
everywhere, and silently. everywhere, and silently.
So `memo wake` pages the document into parts that fit all of them So `memo wake` pages the document into parts that fit all of them
@ -174,7 +174,7 @@ $MEMORY_DIR/
`memo` and are the only home for them. `memo` and are the only home for them.
ENTRY_CHARS=280 longest a memory may be ENTRY_CHARS=280 longest a memory may be
WAKE_LINES=256 how many lines `memo wake` prints (~16k tokens) WAKE_LINES=208 how many lines `memo wake` prints (~16k tokens)
PART_CHARS=20000 how much of it fits in one command's output PART_CHARS=20000 how much of it fits in one command's output
PART_LINES=500 ...and in how many lines PART_LINES=500 ...and in how many lines
``` ```

2
memo
View file

@ -21,7 +21,7 @@ sys.path.insert(0, os.path.dirname(os.path.realpath(__file__)))
from blocks import cover # noqa: E402 from blocks import cover # noqa: E402
ENTRY_CHARS = 280 ENTRY_CHARS = 280
WAKE_LINES = 256 WAKE_LINES = 208 # ~16k tokens of dense text, in 3 parts
RAW_MAX = 16 # blocks up to this many memories compress from the raw log RAW_MAX = 16 # blocks up to this many memories compress from the raw log
# Every harness truncates a command that prints too much, and each drops a # Every harness truncates a command that prints too much, and each drops a

4
test.py
View file

@ -27,8 +27,8 @@ DEFAULTS = {k: getattr(cli, k) for k in
("ENTRY_CHARS", "WAKE_LINES", "PART_CHARS", "PART_LINES")} ("ENTRY_CHARS", "WAKE_LINES", "PART_CHARS", "PART_LINES")}
N = 2000 N = 2000
WAKE_LINES = 256 WAKE_LINES = cli.WAKE_LINES # the shipped budget, not a second copy of it
PART_CHARS = 20000 PART_CHARS = cli.PART_CHARS
# Verified caps of the harnesses in the wild: Claude Code cuts a command's # Verified caps of the harnesses in the wild: Claude Code cuts a command's
# output at 30,000 chars (middle), pi at 50 KB / 2000 lines (head), Codex # output at 30,000 chars (middle), pi at 50 KB / 2000 lines (head), Codex
# budgets 10,000 tokens. A part must fit the strictest of each kind. # budgets 10,000 tokens. A part must fit the strictest of each kind.