feat: add AtMem memory provider - #43

Summary

Adds AtMem as a local provider for Agent Memory Benchmark.

The adapter:

• stores each benchmark document in a user-scoped local SQLite memory;
• uses AtMem's deterministic extraction and lifecycle handling during ingestion;
• combines lexical and graph retrieval with the released calibrated direct-support decision, so candidate rank alone cannot inject unrelated memory;
• maps returned records back to the benchmark source document IDs;
• returns AtMem's retrieval and decision evidence in raw_response; and
• requires no API key or external service.

AtMem is pinned to 2.2.7b1 so the benchmark lock resolves the exact published package and its stable atmem-atbot==0.1.0 companion.

Reproduce

uv run amb run \
  --dataset precisionmembench \
  --split single-turn \
  --memory atmem \
  --mode retrieval

Validation

• Targeted adapter tests against the installed atmem==2.2.7b1: 2 passed.
• Ruff check and format check passed for the new provider and tests.
• uv lock --check and git diff --check passed.
• A local run over all 77 PrecisionMemBench cases produced 38/77 total passes, 18/43 active passes, mean precision 0.3534, and mean recall 0.5233. This is reported as a local smoke result; no leaderboard output is included in this PR.

The repository's complete environment cannot be installed on this macOS x86_64 host because the existing lancedb==0.29.2 dependency has no compatible wheel. The AtMem adapter itself was tested in an isolated Python 3.12 environment using the published wheel.

AI assistance

Codex generated most of the adapter and tests. The public AtMem APIs were checked against the tagged release, and the tests and benchmark smoke above were run against the published wheel.