A summary is not a causal model.

Video caption

A summary is not a causal model. A later edit removes a guard to simplify the flow. Can a fresh worker explain and preserve the invariant? #EricFieldNotes

Full written post / accessible read

Persisted memory helps an agent resume. But a summary can preserve that a guard exists while losing the incident that made it necessary. The next context window then sees an awkward check and may optimize it away without understanding the failure it prevents.

Imagine a duplicate event was once possible only during an obscure partner retry. The original engineer added a ledger check. Months later, a fresh agent sees two checks and deletes one. Its tests pass because the retry fixture did not travel with the handoff.

For each consequential invariant, retain the incident or counterexample, the accepted behavior and a test that fails when the guard is removed. Link that packet to the code path. Then a new agent can retrieve evidence without loading the entire old conversation.

Start a fresh session with only the durable packet. Ask it to modify the path, predict the protected retry outcome, and run the negative control. Do this because a memory system is useful only if the next worker finds the right causal fact and the gate survives its patch.

#EricFieldNotes

X thread draft

Video belongs on the first post; the rest gives the full argument. Check the native composer before sending.

A summary is not a causal model. A later edit removes a guard to simplify the flow. Can a fresh worker explain and preserve the invariant? #EricFieldNotes

Persisted memory helps an agent resume. But a summary can preserve that a guard exists while losing the incident that made it necessary. The next context window then sees an awkward check and may optimize it away without understanding the failure it prevents.

Imagine a duplicate event was once possible only during an obscure partner retry. The original engineer added a ledger check. Months later, a fresh agent sees two checks and deletes one. Its tests pass because the retry fixture did not travel with the handoff.

For each consequential invariant, retain the incident or counterexample, the accepted behavior and a test that fails when the guard is removed. Link that packet to the code path. Then a new agent can retrieve evidence without loading the entire old conversation.

Start a fresh session with only the durable packet. Ask it to modify the path, predict the protected retry outcome, and run the negative control. Do this because a memory system is useful only if the next worker finds the right causal fact and the gate survives its patch.

Evidence and boundary

On-screen label: ILLUSTRATIVE HANDOFF. Illustrative cases are not measured incidents. Research papers and vendor documents support the stated mechanism only within their studied or documented scope.

• OpenAI Agents SDK sessions (S53)
• LangGraph persistence (S54)
• Liu et al., Lost in the Middle (S80)