How the Agent Memory Advisor decides.

Included and excluded

This is a representative v0.1 set chosen to cover different architecture families: developer continuity, application memory, temporal graphs, enterprise knowledge, reflective learning, stateful runtimes, and the no-dedicated-memory outcome.

It is not a market-size ranking, a popularity list, an affiliate comparison, or a claim that the listed systems are interchangeable. Products with unclear identity, licensing, or public evidence remain in the catalog only when the uncertainty itself is useful to a decision.

Decision procedure

1. Determine the workload and memory shape.
2. Make canonical-state ownership and portability explicit.
3. Evaluate every selected hard requirement as PASS, FAIL, or UNKNOWN. FAIL eliminates; UNKNOWN remains visible only as conditional or validation-required.
4. Rank fail-free candidates by architecture family, hard-certainty tier, explicit preferences, and evidence strength.
5. Return one of four generated states: strong recommendation, validation required, two-system pilot, or constraint conflict.
6. Return a first system to evaluate, alternatives, why-not explanations, uncertainty, and a five-test pilot.

Evidence meanings

AMSB tested

Tested in the Agent Memory Sovereignty Bench under a pinned configuration.

Independent research

Supported by a paper or evaluation outside the product team.

Implementation verified

The current official repository or documentation establishes the capability.

Vendor benchmark

A project-published performance claim, not an independent finding.

Architectural inference

A reasoned interpretation of the architecture, not a tested guarantee.

Not tested

No evidence is claimed for this behavior in the current snapshot.