Improve the search memories in KeyValueMemory Toolset

[Pooria90]

Summary

Closes #1249

Adds pluggable relevance ranking to the prebuilt key-value memory toolset. The search_memories tool previously did a plain case-insensitive substring match; it now ranks results through a swappable SearchAlgorithm. Two implementations ship:

• LexicalSearch (the new zero-dependency default; key-weighted token overlap, exact-key jackpot, multi-term coverage, and a difflib fuzzy fallback for typos).

• SemanticSearch (opt-in embedding-similarity search that reuses the existing Embedding + VectorBackend infra and keeps a lazily-maintained vector index in sync as memories change).

The whole search layer lives under prebuilt/tools/memory/search/. The KeyValueMemoryToolSet gains a search= argument to select or customize the ranker.

Type of change

• Bug fix
• Feature
• Breaking change
• Docs
• Refactor / chore / build / tests

Checklist

• Lint & format pass (ruff check . && ruff format .)
• Tests added/updated and pass locally (pytest tests)
• Docs updated if user-facing behavior changed
• Breaking changes include migration notes

Notes

Behavior change (non-breaking): search_memories returns the same "- key: value" format and same tool signature, but results are now relevance-ranked rather than raw substring matches. Zero-config usage is unchanged; the default LexicalSearch needs no dependencies or setup.

Swap in semantic search:

from railtracks.prebuilt.tools.memory import SemanticSearch
from railtracks.retrieval.embedding import OpenAIEmbedding
from railtracks.retrieval.stores.vector.backends import InMemoryBackend

memory = rt.prebuilt.KeyValueMemoryToolSet(
    search=SemanticSearch(
        embedding=OpenAIEmbedding(),
        backend=InMemoryBackend(snapshot_path="memory_vectors.json"),  # optional persistence
    ),
)

[Amir-R25]

Looks very cool, two nits but LGTM 🚀

[soulFood5632]

Great work Pooria

[soulFood5632]

LGTM