fix(intelligence): soft mark forgotten memories

[wayyoungboy]

Summary

This PR completes the intelligent-forgetting soft-mark behavior introduced for Memory.get() in #282.

When EbbinghausIntelligencePlugin returns delete_flag=True, memories should be marked as forgotten instead of being physically removed from storage. The sync Memory.get() path already does this, but Memory.search(), AsyncMemory.get(), and AsyncMemory.search() still called delete_memory / delete_memory_async.

Changes

• Add a forget-marker payload helper for sync and async memory modules.
• Update Memory.search() to call update_memory() with should_forget and marked_for_forgetting_at instead of deleting.
• Update AsyncMemory.search() to call update_memory_async() instead of deleting.
• Update AsyncMemory.get() to soft mark and return the current result, matching Memory.get() behavior.
• Add regression tests covering sync search, async search, and async get soft-mark behavior.

Test plan

• pytest tests/unit/test_intelligence_forget_soft_mark.py -v
• pytest tests/unit/intelligence -v
• pytest tests/unit/test_memory.py -v
• git diff --check

[lightzt99]

@wayyoungboy The four-path alignment in this PR is correct. While reviewing, we traced the full lifecycle of the forget marker and found a broader architectural gap worth discussing.

What this PR fixes

Before #282, all four intelligence-plugin paths hard-deleted memories. #282 fixed only Memory.get(). This PR fixes the remaining three — Memory.search(), AsyncMemory.get(), and AsyncMemory.search() — so all four paths now consistently soft-mark rather than physically delete.

The write/read gap

When a memory is soft-marked, two fields are written to storage:

should_forget = True
marked_for_forgetting_at = <timestamp>

However, a full-text search of the codebase shows these fields are written but never read by any component:

Component	Reads should_forget?	Effect
Storage search/list/get	❌	Forgotten memories returned as normal
API response transform	❌	Fields dropped before reaching consumers
Plugin on_get/on_search	❌	Recalculates should_forget() from scratch each time
GC / cleanup job	❌	Does not exist

The same gap exists for review_schedule, which process_memory_metadata() generates on every memory access but get_review_schedule() has zero external callers.

The Ebbinghaus lifecycle: four signals, two consumers

Signal              Written by              Consumed by
─────────────────────────────────────────────────────────
review_schedule     process_memory_metadata()   ❌ none
should_forget       on_get / on_search          ❌ none
should_promote      process_memory_metadata()   ✅ IntelligentMemoryManager
should_archive      process_memory_metadata()   ✅ IntelligentMemoryManager

The promotion/archival signals have a consumer (optimize_memories() manages in-memory tiers). The forget and review signals produce data that no component acts on.

Suggested next steps

The soft-mark fix in this PR is a necessary first step, but the forget mechanism becomes meaningful only when a reader exists. Two natural follow-ups:

1. Filter layer: storage queries should exclude should_forget=True memories by default, with an opt-in flag (include_forgotten=True) for audit/recovery scenarios.
2. GC job: periodically hard-delete memories where marked_for_forgetting_at exceeds a configurable retention window (e.g., 30 days), giving a two-stage safety net: soft-mark → grace period → hard delete.

These are outside the scope of this PR but worth tracking as follow-up issues.

[wayyoungboy]

Review update: no blocking issues found in this PR.

I agree with the new comment that should_forget / marked_for_forgetting_at still need a reader path to become fully effective, such as default filtering plus an opt-in include_forgotten mode, and eventually a retention/GC job. That is a real follow-up gap, but I would not block this PR on it: this change fixes the narrower inconsistency where Memory.search(), AsyncMemory.get(), and AsyncMemory.search() still hard-deleted while Memory.get() had already moved to soft marking.

Local verification:

PYTHONPATH=src pytest tests/unit/test_intelligence_forget_soft_mark.py -q
# 3 passed

git diff --check
# passed

[lightzt99]

Thanks @wayyoungboy for the thorough analysis. Agreed — this PR closes the consistency gap, and the filter/GC follow-up is tracked separately. Merging.