Keep state outside the model

Plans, findings, and file lists live in notes outside the model. The model context holds only the current working slice.

External notes survive summarisation and restarts. Reviewers can read the same facts the run used.

Context growth arithmetic

Each loop step appends observations to the context. Ten steps of 800 tokens on a 2000-token start reach 10000 tokens, as the calc block shows.

Token counts are approximate and tokenizer-dependent. Common tokenizers use about 3 to 4 characters per token for English prose, with fewer characters per token for code, numbers, hex, and non-Latin text.

In the basic form, attention cost grows with the square of the sequence length, and efficient variants exist. Long-context use is less reliable than short-context use, and models tend to use information at the start and end of long inputs better than the middle, a tendency reported in studies rather than a law.

1
2
3
4
5

start = 2000 tokens, appended per step = 800 tokens, steps = 10 (stated example values)
total = 2000 + 10 x 800 = 2000 + 8000 = 10000 tokens
summary = 1500 tokens = 1500 / 10000 = 0.15 = 15%
saving = 10000 - 1500 = 8500 tokens = 85%
check: 15% + 85% = 100%

Summarise without losing facts

A summary keeps decisions, open items, and file paths. It drops raw logs and keeps pointers to them.

Summarise on a fixed cadence or size threshold. What survives a summary varies by model: measure it on the task.

Common mistakes

Related topics

• Task decomposition

• Budgets and enforcement

• Cost arithmetic