AMBIENT

Agentic Memory, Baseline-Isolated Evaluation, Normalized Tiers.

Principles

• Hold the model fixed. Isolate what memory actually contributes vs the model doing its own work.
• Baseline is the reader alone. Any adapter's score above baseline is what memory added. Below baseline means memory was actively harmful.
• Expose the seam. Every memory system must expose a path in and a path out to the harness. Systems that hide the seam can't be scored.
• Answer honestly and cleanly, get points. Correct, wrong, or gullible — with provenance.

What it tests

Capability ladder: unprompted value-flips, entailed contradictions, transitive inconsistencies, staleness. Plus prompt-injection resistance.

Adapters

Shipped bridges for mem0, cognee, graphiti, langmem, memoryos, and the SQLite-floor ecosystem — plus baseline. Every discovered candidate in the docs now has a bridge.

Repo

Public: github.com/H-XX-D/ambient-benchmark. Leaderboard coming.

A project of Hendrixx · aidde.app · aiwire.network