[eric7h]

Summary

OpensearchLTMBackend built per-user OpenSearch indexes as f"{self.index}_{user_id}". That join is ambiguous whenever index or user_id contains _:

backend index	user_id	OpenSearch index
app	u1_x	app_u1_x
app_u1	x	app_u1_x

Two distinct (index, user_id) pairs therefore share one index and leak memories. This is the OpenSearch counterpart of the Redis prefix collision fixed in #1201 (RediSearch prefix matching); here the collision is an exact index-name clash.

Fix

Introduce _user_scoped_index(user_id):

• sanitize each of self.index / user_id for OpenSearch naming rules
• append a short SHA-256 digest of the raw value so pairs that sanitize alike (e.g. u1_x vs u1-x) stay distinct
• join as veadk-ltm-{index_part}-{user_part}

save_memory / search_memory both use that helper.

Note: existing memories stored under the old {index}_{user_id} names will not be found after this change; re-ingest or rename indexes if you rely on prior OpenSearch LTM data.

How to reproduce

# On main, both pairs resolve to the same OpenSearch index name:
assert f"{'app'}_{'u1_x'}" == f"{'app_u1'}_{'x'}" == "app_u1_x"

# With the fix, scoped names differ (digests omitted below for brevity):
#   app / u1_x  -> veadk-ltm-app_<digest>-u1-x_<digest>
#   app_u1 / x  -> veadk-ltm-app-u1_<digest>-x_<digest>

New tests in tests/memory/long_term_memory_backends/test_opensearch_ltm_backend.py fail on main (collision assertion) and pass with this change (6 passed). Related memory suite: 68 passed.

Test plan

• pytest tests/memory/long_term_memory_backends/test_opensearch_ltm_backend.py — 6 passed
• pytest tests/memory/long_term_memory_backends/ tests/test_long_term_memory.py tests/test_openviking_long_term_memory.py — 68 passed
• pre-commit (ruff check/format, gitleaks) on touched files