fix(memory): guard against duplicate-entry trap

[AIalliAI]

Cherry-pick of PR NousResearch#60090 by Vissirexa.

Fixes: NousResearch#60089
Pipeline: TRIAGE DONE | INTEGRATOR DONE | VERIFY PASS | REVIEWER PENDING | SECURITY PENDING | OPSEC PENDING

[AIalliAI]

Hermes Agent Code Review — VERDICT: APPROVED ✅

PR: #5 — cherry-pick of NousResearch#60090 by Vissirexa
Files: tools/memory_tool.py (+37), tests/tools/test_memory_tool.py (+43)

---

What This Fixes

When MemoryStore.replace() hits an ambiguous match (old_text matches multiple entries with different content), the model gets "Multiple entries matched" → falls back to add() → creates a duplicate of an entry that already says what the model wanted to write. This is a repeatable tool-misuse trap.

What Changed

Two call sites, both exact-string-identity checks, scoped to action="replace":

1. MemoryStore.replace() (line 429): When multiple unique entries match AND the replacement text already exists verbatim as a separate entry, return a targeted error with actionable guidance instead of the generic ambiguous-match message. Includes matches: previews for visibility.

2. memory_tool() dispatcher (line 1038): Same pre-validation runs before the approval gate for earlier, equally targeted feedback.

✅ Correctness

• The guard only fires when len(unique_texts) > 1 AND new_content in entries — exactly the trap scenario
• When replacement text is genuinely new, the original generic error is preserved (verified by test)
• Edge cases: identical-duplicate matches fall through (safe to operate on first), single match falls through (normal replace path)
• Edge case where new_content equals one of the matched entries: guard fires correctly, guiding to batch operations
• Comparison uses stripped strings consistently on both sides
• Dispatcher check is read-only, no lock needed; benign TOCTOU — store-level catches any race under lock

✅ Tests

All 86 tests pass (83 existing + 3 new). Adjacent files 4 pass.

New tests:

• test_replace_ambiguous_match_with_duplicate_replacement_gets_targeted_hint — store guard fires with targeted message ✅
• test_replace_ambiguous_match_without_duplicate_keeps_generic_error — regression: original error unchanged ✅
• test_replace_dispatcher_catches_duplicate_trap_before_store — dispatcher guard fires ✅

💡 Suggestion (non-blocking)

The dispatcher pre-validation (line 1038) returns a targeted error but does not include the matches key with previews of the matched entries — unlike the store-level guard which includes matches: previews. For consistency and richer model context, consider passing matched entries through store._previews() and including them here too.

Verdict: APPROVED ✅

The fix is correct, well-targeted, non-invasive, and does exactly what it claims. No regressions, good test coverage. Ready to merge.