[santoshkumarradha]

A stalled background memory request could consume all 45 seconds, leaving its fallback no time and dropping the exchange. Reserve half the remaining deadline for the configured small-work model, then retry the original exchange once after two seconds if the pass still fails transiently. The chat stays responsive and successful recovery saves normally without a failure notice.

Cancellation stops recovery; permanent failures do not cause repeated requests. Exhausted recovery retains the technical journal record and explains the direct /remember path. Pending exchanges remain owned by the open chat and are not persisted across restarts. The system prompt now acknowledges automatic capture instead of incorrectly claiming /remember is required.

Proof

The real terminal acceptance test TestMemoryRecoversFromTimeoutOnTheRealTerminal passed on 7a6ed127d3e5e70a53f73ad524ad323cb8c6fd9c in 106.58 seconds using OpenRouter deepseek/deepseek-v4.1-flash for every model role. All data below are fictional:

1. Two injected memory-request stalls exhausted the first 45-second pass while the chat answered normally.
2. A later background attempt saved the original weekend-envelope convention exactly once, without another user message.
3. A fresh chat recalled it. A natural-language correction replaced violet with copper envelopes.
4. Another fresh chat used the corrected convention to create packaging.json with {"weekend_envelope_color":"copper"}.
5. All 26 observed requests and 51 model-log receipt rows named the exact requested model. Recovered saves displayed no terminal failure notice.

Reproduce after make build with an OpenRouter key configured:

CODEAF_E2E_MODEL=deepseek/deepseek-v4.1-flash go test -tags e2e ./internal/e2e -run '^TestMemoryRecoversFromTimeoutOnTheRealTerminal$' -count=1 -timeout 8m -v

Focused regressions also cover the exact 22.5-second primary / 45-second overall deadline split, same-model recovery, cancellation, bounded exhaustion, technical journal retention, direct manual saving after exhaustion, and session-close lifecycle. Reflex tests pass under the race detector. make pr-ready passed: build, vet, manual gates, structural laws and all affected-package tests (including all eight session shards). All seven GitHub check/status results passed on the tested head: PR gate.