fix(memory): query prompt and seed metadata whose keys are not SQL identifiers

[Chen Yufeiyang (feiiiiii5)]

Fixes the free-form metadata keys that no backend can filter on.

Description

MemoryInterface documents prompt_metadata and seed metadata as "a free-form dictionary for tagging prompts with custom metadata". Both SQL backends spelled each key into the JSON path and reused it as the bind-parameter name:

json_conditions = " AND ".join(
    [f"JSON_EXTRACT(prompt_metadata, '$.{key}') = :{key}" for key in prompt_metadata]
)
condition = text(json_conditions).bindparams(**dict(prompt_metadata.items()))

Only [A-Za-z0-9_] survives that. Every other character raises before the query runs:

await memory.get_message_pieces_async(prompt_metadata={"attack.id": "v"})
# sqlalchemy.exc.ArgumentError: This text() construct doesn't define a
# bound parameter named 'attack.id'

So the filter cannot be expressed at all for a dotted, hyphenated or spaced key — and a dot was additionally read as a path step, meaning a key that did survive was looking up a nested object. The existing coverage uses only identifier-safe keys (finish_reason, status, incomplete_reason), which is why this went unnoticed.

The same two code paths serve seed metadata (_get_seed_metadata_conditions), so seed filtering has the same problem.

Changes

• sqlite_memory.py: _get_message_pieces_prompt_metadata_conditions and _get_seed_metadata_conditions now bind the JSON path with the key quoted into it ('$."{key}"') and give each condition positional parameter names.
• azure_sql_memory.py: the same in _get_metadata_conditions, which backs both prompt and seed metadata there.

This is how #3063 fixed the identical interpolation in the label conditions, so the three paths now agree.

JSON_EXTRACT(prompt_metadata, :pm_path_0) = :pm_value_0   --  pm_path_0 = '$."attack.id"'
JSON_VALUE(prompt_metadata, :md_path_0) = :md_value_0    --  Azure SQL

Binding the path also keeps keys and values out of the statement text entirely.

Tests

tests/unit/memory/memory_interface/test_interface_prompts.py queries four keys holding a dot, hyphen, space and colon, and one query naming three keys at once. Against this branch's parent they fail with the ArgumentError above. The plain-key cases the suite already covered still pass.

No documentation change: MemoryInterface already documents these dictionaries as free-form, and this PR makes that true rather than narrowing it.

Tests and Documentation

pytest tests/unit/memory/memory_interface/ tests/unit/memory/test_sqlite_memory.py \
       tests/unit/memory/test_azure_sql_memory.py -q
# 648 passed, 1 skipped
ruff check   # All checks passed!
ruff format --check   # 3 files already formatted

JupyText not run: no code samples or documentation changed.

Not addressed

A key containing a double quote cannot be expressed in a SQLite or T-SQL JSON path, so after this change it matches nothing rather than raising. That is the same residual #3063 left in the label path; closing it would need a different lookup strategy than JSON paths, which is more than this fix should take on.

🤖 Generated with Claude Code