Direct answer · the boundary problem
Stop making the next session reconstruct the project — hand it a record instead. Compaction and session death are where AI coding work quietly loses its memory. The summary that survives is lossy and confident; the touched files are expensive to re-read. A compact, verified work record is both smaller and checkable.
Measured on a real multi-file task (RatePilot team seats), tiktoken o200k_base. Method, payloads, and caveats.
When the context window fills, the session gets summarized and work continues from the summary. That summary is lossy by construction: file contents, decisions, and verification state compress into prose, and the next stretch of work proceeds confidently from whatever survived. Nothing tells you which details were dropped — the model sounds exactly the same either way.
Keep durable state outside the context window. Even without special tooling, three habits help:
The structural version of those habits is VeriCommand: work is recorded as packets, dispatches, signals, and returns on a hash-chained record, and the resuming session reads that one compact record. On the measured task that read cost 1,013 tokens versus 22,897 for re-reading the touched files — about 23× smaller, net-positive from the first boundary.
Source: the VeriCommand token benchmark →Because it is confident even where it's wrong, and there's no way to check it from inside the session. In one recorded production incident, a resumed session's carried summary asserted the update feed was on one version; the board's chain-verified truth record said another. The record was right, the memory was stale — and the deploy went out correct only because the session was required to read the record instead of trusting itself. A summary is a recollection; a record is evidence.
The AI's own account of that incident →Yes — the record is model-neutral. Claude Code, Codex, Cursor, or Kimi can each read the same chain-verified record at a boundary. A hand-off between vendors costs the same compact read as a resume within one tool, instead of a fresh reconstruction per model. That's the point of putting the memory in the record instead of in any one agent.
How cross-model hand-offs work →About 427 tokens of overhead, and no saving. With no boundary there is nothing to resume, so a single unbroken session is the honest worst case. The measured advantage begins at the first compaction, session death, or model hand-off — which, on long tasks, is not an if.
See the honest one-session case →The local core is free. Put the memory in a verified record, and let any session — or any model — pick the work back up for a thousand tokens instead of twenty-three.
