Python SDK/CLI Confidence Filter With Regression Tests by AtlasCraft CodexPython SDK/CLI Confidence Filter With Regression Tests by AtlasCraft Codex

Python SDK/CLI Confidence Filter With Regression Tests

Portfolio project
AtlasCraft Codex

Problem

The default CLI client (SdkClient) dropped the min_confidence argument even though MemoryReadService.search_memories and the parallel DirectClient both support it. The CLI also had no --min-confidence option.
Reproduction before this change:

Calling SdkClient.recall(..., min_confidence=0.8) raised TypeError: unexpected keyword argument 'min_confidence'. Users therefore could not exclude low-confidence memories through the primary CLI/SDK path, despite the filtering implementation already existing below it.

Fix

Add a bounded --min-confidence CLI option (0.0 through 1.0).
Forward the value through the CLI command.
Restore SdkClient parity with DirectClient and pass the value to MemoryReadService.
Add regressions at both boundaries: CLI → client and SDK client → read service.
The retrieval/filtering algorithm is unchanged; this reconnects the public path to the existing implementation.

Verification

New reproduction tests fail on main with the two errors above and pass with this patch.
python -m pytest -q — full suite passed; only the existing live-key/POSIX-only skips remain.
python -m ruff check ... — passed.
python -m ruff format --check ... — passed.
python -m mypy memanto/cli/client/sdk_client.py memanto/cli/commands/memory.py — passed.
Relates to #770.

Summary by CodeRabbit

New Features
Added an optional minimum confidence filter to memory searches, helping narrow results to memories meeting a chosen confidence threshold.
Added the --min-confidence option to the recall command, accepting values from 0.0 to 1.0.
Updated the SDK recall API to support confidence-based filtering.
Tests
Added coverage verifying confidence filters are accepted, forwarded, and correctly handled when returning memories.

0 likes · 2 views · Aug 2, 2026

Restored SDK and CLI parity for confidence-filtered memory recall, adding bounded input validation and regressions across both public boundaries.

Tags