Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@nowledge-mem-openclaw-plugin/src/config.js`:
- Around line 52-56: The captureMinInterval value currently only clamps a lower
bound; update the parsing logic for captureMinInterval (in the config handling
where captureMinInterval is computed) to also enforce the schema's upper bound
by applying Math.min(86400, ...) around the truncated value similar to how
maxRecallResults uses Math.min/Math.max; ensure you still use Math.trunc and
Math.max(0, ...) but wrap that result with Math.min(86400, ...) so values above
86400 are capped to the manifest limit.

In `@nowledge-mem-openclaw-plugin/src/hooks/capture.js`:
- Line 264: The cooldown calculation incorrectly treats an explicit 0 as falsy;
in buildAgentEndCaptureHandler replace the fallback logic so cooldownMs is
derived directly from cfg.captureMinInterval (e.g., remove the "|| 300"
fallback) — i.e., set cooldownMs = cfg.captureMinInterval * 1000 so a configured
0 yields 0ms and the existing if (cooldownMs > 0) guard works as intended;
reference symbols: cfg.captureMinInterval, buildAgentEndCaptureHandler,
cooldownMs.
- Around line 271-280: The cooldown early-return is dropping legitimate
messages; remove the cooldown gate that returns before calling
appendOrCreateThread (the block using cooldownMs, buildStableThreadId, and
_lastCaptureAt) and instead apply the cooldown only around the expensive
triage/distill path after appendOrCreateThread and after you inspect
messagesAdded; call appendOrCreateThread unconditionally so dedup/zero-delta
logic (messagesAdded === 0) can skip work, then check cooldown and update
_lastCaptureAt before running triage/distill so only that expensive step is
rate-limited.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@nowledge-mem-openclaw-plugin/src/hooks/capture.js`:
- Around line 291-303: The code is recording cooldown too early by calling
_lastCaptureAt.set(result.threadId, Date.now()) before the short-conversation
filters run, which can wrongly suppress the first eligible triage; update the
logic in the capture flow so the cooldown timestamp for result.threadId is only
set after all triage eligibility checks (including the short-conversation
filters around captureMinInterval) pass and just before/after you perform the
triage/distillation action, i.e., remove or move the _lastCaptureAt.set call out
of the early cooldown block and into the code path that actually executes triage
so that you only record a capture when triage runs.

---

Outside diff comments:
In `@nowledge-mem-openclaw-plugin/src/hooks/capture.js`:
- Around line 272-284: The appendOrCreateThread call (result = await
appendOrCreateThread({...})) is executed before checking cfg.autoCapture,
causing thread writes even when autoCapture is false; update the agent_end
handler so you first check if cfg.autoCapture is truthy and only then call
appendOrCreateThread (or skip and return early), i.e., move or guard the
appendOrCreateThread invocation behind the existing if (!cfg.autoCapture)
return; using the same parameters (client, logger, event, ctx, reason:
"agent_end") to ensure thread append and subsequent triage/distill run only when
autoCapture is enabled.

---

Nitpick comments:
In `@nowledge-mem-openclaw-plugin/src/hooks/capture.js`:
- Around line 9-12: The Map _lastCaptureAt in capture.js can grow unbounded;
update the capture logic that reads/writes _lastCaptureAt to evict stale entries
by age (e.g., remove entries older than a configured TTL like 1 hour) or cap its
size with an LRU eviction policy; implement the eviction inside the functions
that call/modify _lastCaptureAt (the per-thread capture cooldown helpers) and/or
run a lightweight periodic cleanup (setInterval) to prune old threadId keys so
long-lived agents do not leak memory.

Verify each finding against the current code and only fix it if needed.

Outside diff comments:
In `@nowledge-mem-openclaw-plugin/src/hooks/capture.js`:
- Around line 269-283: The call to appendOrCreateThread in the agent_end handler
runs unconditionally and can persist messages even when cfg.autoCapture is
false; wrap or move that call so it only executes when cfg.autoCapture is true
(i.e., check cfg.autoCapture before invoking appendOrCreateThread or place the
appendOrCreateThread block inside the existing if (!cfg.autoCapture) guard),
ensuring all logic that appends or creates threads is gated by cfg.autoCapture
while preserving any other unconditional behaviour in agent_end.

---

Nitpick comments:
In `@nowledge-mem-openclaw-plugin/src/hooks/capture.js`:
- Around line 9-11: _current code creates a process-global Map named
_lastCaptureAt that never evicts keys; modify the capture logic where
_lastCaptureAt is written and read (the code paths that set/check per-thread
cooldown) to perform opportunistic cleanup: when setting an entry, prune keys
older than a configurable TTL (e.g., compare stored timestamp to Date.now() -
TTL) and/or enforce a max size by removing least-recently-used keys; ensure the
TTL and max-size constants are declared near _lastCaptureAt and that reads also
ignore/cleanup expired entries so stale thread IDs do not accumulate.