Summary
A single long-lived session can grow an extract far larger than Haiku's context window. save-session.sh embeds the entire extract into the Step-4 prompt with no size cap, so the claude -p --model haiku call fails (input exceeds context length), set -e aborts the save, and the daily rotation (now.md → today-YYYY-MM-DD.md) silently stops. It is self-reinforcing: a failed save never advances the saved position, so the same session re-extracts the full transcript on every subsequent save and fails identically — until an unrelated small session happens to run.
Confirmed on marketplace remember@claude-plugins-official v0.7.3, Windows / Git Bash.
This is related to but distinct from #94. #94 attributes FAILED at line ~179 to machine-load-induced hangs/timeouts plus the dead set -e error handler. My incident is not a hang — every failing save logs an oversized extract and fails in ~25s, not a 180s timeout. The two are complementary: a size cap prevents the failure; #94's || HAIKU_EXIT=$? + timeout change makes any residual failure visible and non-blocking. Neither alone is sufficient.
Evidence (from .remember/logs/memory-*.log)
Daily rotation went dark for two full days. Every save against the offending session logged:
[extract] 5389 exchanges (292 human)
[haiku] calling (branch: ...)
[rotate] ERROR: tar failed for 30 logs # non-fatal red herring (log.sh:220 else-branch returns 0)
[error] FAILED at line 175 (exit 1)
Meanwhile small sessions on the same machine saved fine, interleaved with the failures:
20:34 [extract] 33 exchanges (2 human) → OK
20:57 [extract] 5389 exchanges (292 human) → FAILED at line 175
20:59 [extract] 51 exchanges (5 human) → OK, [write] appended
5389 exchanges is well past Haiku's 200K-token window, so the CLI rejects the input and exits non-zero. Because the failing line is the HAIKU_JSON=$(... < "$TMP_PROMPT") command substitution, set -e trips there and the generic ERR trap logs FAILED at line 175 — the helpful if [ "$HAIKU_EXIT" -ne 0 ] diagnostic never runs (the dead-handler issue, already noted in #94), which is why the real "context length exceeded" error is invisible in the logs.
Root cause
scripts/save-session.sh Step 3 calls pipeline.shell build-prompt "$EXTRACT_FILE" ..., which substitutes the full $EXTRACT_FILE into the prompt {{EXTRACT}} placeholder. There is no upper bound on $EXTRACT_FILE size anywhere between extract (Step 1) and the Haiku call (Step 4). A session that accumulates thousands of exchanges therefore produces a prompt that cannot fit in the model's context.
Suggested fix
Cap the extract (keeping the most-recent tail) before building the prompt. I've been running this locally with no downside — daily rotation resumed immediately:
# After the dry-run block, before "Step 2: Get last entry":
EXTRACT_MAX_BYTES=$(config ".thresholds.extract_max_bytes" 300000)
if [ -f "$EXTRACT_FILE" ]; then
EXTRACT_BYTES=$(wc -c < "$EXTRACT_FILE" | tr -d ' ')
if [ "${EXTRACT_BYTES:-0}" -gt "$EXTRACT_MAX_BYTES" ]; then
TMP_CAPPED=$(mktemp "${TMPDIR:-/tmp}"/remember-extract-capped-XXXXXX)
printf '[NOTE: transcript truncated to the last %s of %s bytes — summarize the most recent work below]\n\n' \
"$EXTRACT_MAX_BYTES" "$EXTRACT_BYTES" > "$TMP_CAPPED"
tail -c "$EXTRACT_MAX_BYTES" "$EXTRACT_FILE" >> "$TMP_CAPPED"
mv "$TMP_CAPPED" "$EXTRACT_FILE"
log "extract" "capped ${EXTRACT_BYTES}b -> ${EXTRACT_MAX_BYTES}b (oversized-prompt guard)"
fi
fiDefault 300 KB (~75K tokens) leaves generous headroom under Haiku's 200K window; configurable via .thresholds.extract_max_bytes. Keeping the tail is appropriate since the summary should reflect the most recent work. (A token-aware cap inside build-prompt would be even better, but the byte cap is a robust, dependency-free floor.)
Combined with #94's timeout + reachable error handler, an oversized or slow Haiku call would then degrade gracefully instead of silently halting all memory writes.
Environment
remember@claude-plugins-official v0.7.3- Windows 10 / Git Bash (MSYS coreutils)
- Symptom window: two days of missing
today-*.md; FAILED at line 175 also present in 8 separate daily logs over the prior weeks whenever a session got large.
Summary
A single long-lived session can grow an extract far larger than Haiku's context window.
save-session.shembeds the entire extract into the Step-4 prompt with no size cap, so theclaude -p --model haikucall fails (input exceeds context length),set -eaborts the save, and the daily rotation (now.md→today-YYYY-MM-DD.md) silently stops. It is self-reinforcing: a failed save never advances the saved position, so the same session re-extracts the full transcript on every subsequent save and fails identically — until an unrelated small session happens to run.Confirmed on marketplace
remember@claude-plugins-officialv0.7.3, Windows / Git Bash.This is related to but distinct from #94. #94 attributes
FAILED at line ~179to machine-load-induced hangs/timeouts plus the deadset -eerror handler. My incident is not a hang — every failing save logs an oversized extract and fails in ~25s, not a 180s timeout. The two are complementary: a size cap prevents the failure; #94's|| HAIKU_EXIT=$?+timeoutchange makes any residual failure visible and non-blocking. Neither alone is sufficient.Evidence (from
.remember/logs/memory-*.log)Daily rotation went dark for two full days. Every save against the offending session logged:
Meanwhile small sessions on the same machine saved fine, interleaved with the failures:
5389 exchanges is well past Haiku's 200K-token window, so the CLI rejects the input and exits non-zero. Because the failing line is the
HAIKU_JSON=$(... < "$TMP_PROMPT")command substitution,set -etrips there and the genericERRtrap logsFAILED at line 175— the helpfulif [ "$HAIKU_EXIT" -ne 0 ]diagnostic never runs (the dead-handler issue, already noted in #94), which is why the real "context length exceeded" error is invisible in the logs.Root cause
scripts/save-session.shStep 3 callspipeline.shell build-prompt "$EXTRACT_FILE" ..., which substitutes the full$EXTRACT_FILEinto the prompt{{EXTRACT}}placeholder. There is no upper bound on$EXTRACT_FILEsize anywhere between extract (Step 1) and the Haiku call (Step 4). A session that accumulates thousands of exchanges therefore produces a prompt that cannot fit in the model's context.Suggested fix
Cap the extract (keeping the most-recent tail) before building the prompt. I've been running this locally with no downside — daily rotation resumed immediately:
Default 300 KB (~75K tokens) leaves generous headroom under Haiku's 200K window; configurable via
.thresholds.extract_max_bytes. Keeping the tail is appropriate since the summary should reflect the most recent work. (A token-aware cap insidebuild-promptwould be even better, but the byte cap is a robust, dependency-free floor.)Combined with #94's
timeout+ reachable error handler, an oversized or slow Haiku call would then degrade gracefully instead of silently halting all memory writes.Environment
remember@claude-plugins-officialv0.7.3today-*.md;FAILED at line 175also present in 8 separate daily logs over the prior weeks whenever a session got large.