Community challenge: try to break Lians temporal recall

[Lians-ai]

The challenge

Facts change. AI memory systems often keep both the old and new versions, then return stale information during recall.

We want developers to independently test whether Lians actually prevents that failure.

Run it

git clone https://github.com/Lians-ai/Lians.git
cd Lians/agentmem
pip install -e ".[dev]"
pytest tests/test_supersession_benchmark.py tests/test_recall_quality.py -v

No API key is required.

What to test

1. Add a fact with an event timestamp.
2. Add a corrected or superseding fact.
3. Confirm present-time recall excludes the stale version.
4. Use point-in-time recall to reconstruct what was known before the correction.
5. Try an adversarial example where wording changes but the underlying entity and metric remain the same.

The repository currently reports:

• 0 stale facts in top-5 recall on the included benchmark
• 100% supersession accuracy on 22 fact pairs
• Point-in-time reconstruction through recall_at and snapshot

We want independent attempts to break those claims.

Share evidence

Reply with:

• The fact pair or dataset you tested
• Your environment
• The command or script used
• Expected result
• Actual result
• Whether you found stale recall, incorrect supersession, or reconstruction failure

Useful failures will become regression tests when reproducible. Contributors who find meaningful edge cases will be credited in the repository and invited to a technical pairing session with the maintainers.

Why this matters

In financial, medical, legal, and other regulated workflows, an agent must distinguish what is true now from what was known when a decision was made. A correction should not rewrite history, and history should not contaminate present-time context.

This is a technical evaluation, not a request for stars. Critical results are especially welcome.

Documentation:

• Benchmark methodology
• Testing guide
• Local quickstart