feat: add valkey memory

[edlng]

Summary

Adds a working ValkeyMemory implementation backed by Valkey Search (HNSW vector index) and switches the optional dependency from the async valkey-glide to the synchronous valkey-glide-sync client, which fits the current runtime's blocking call patterns.

Issue

N/A

Changes

• runtime/node/agent/memory/valkey_memory.py — new file implementing ValkeyMemory. Stores each memory item as a Valkey Hash with content, embedding (float32), agent role, and timestamp. Creates an FT index with HNSW COSINE metric at init. Retrieval uses KNN search and converts cosine distance to similarity for threshold filtering.
• entity/configs/node/memory.py — added optional username field to ValkeyMemoryConfig for ACL-based auth; wired it through from_dict and field_specs.
• pyproject.toml — changed optional dep from valkey-glide>=2.4 to valkey-glide-sync>=2.4.
• tests/test_valkey_memory.py — replaced the skip-always integration skeleton with unit tests covering instantiation, update, retrieve (threshold, ordering, errors), count, load/save no-ops, and the lazy import error path. All tests use mocked glide_sync.
• uv.lock — lock file updated for valkey-glide-sync 2.4.1.

Tests

Unit tests in tests/test_valkey_memory.py cover:

• Index creation (happy path, already-exists, missing Search module)
• update() hash storage, float32 encoding, TTL expire calls, empty-input short circuit
• retrieve() result parsing, similarity threshold filtering, descending sort, empty query, search errors
• count_memories() via ft.info
• _get_glide_sync() ImportError when the package is missing

Run with pytest tests/test_valkey_memory.py -v.

[atao2004]

Hey the code looks pretty good I didn't find any bugs. The implementation follows the same patterns as Mem0Memory and GLIDE API usage seems correct.

Just two things; wondering if we need to filter by agent_role as per the spec and to check if it would be helpful to use tls.

[atao2004]

LGTM!