Memory readiness means your agent can recall useful facts without becoming a black box. Score the memory layer on inspectability, supersession, forgetting, portability, source breadth, and tenant isolation; then share the URL with the exact answers preserved for a technical review or vendor shortlist.
Answer each dimension with yes, partial, or no. The score updates in the browser and the result URL preserves only those six answers, so it can be shared without creating an account or storing personal data.
This is a readiness rubric, not a benchmark. Treat the score as a way to decide which control should block a pilot, procurement review, or migration plan.
The agent may remember things, but the memory layer is not reviewable enough for production decisions. Start by adding inspection, scope, and deletion evidence before broadening sources.
Pick one production workflow and instrument the recall path before adding more memories.
The foundation is usable, but one or two weak dimensions can still break trust. Buyers should treat the missing controls as explicit evaluation tasks, not roadmap assumptions.
Turn every partial answer into a concrete acceptance test for your shortlist.
The core control plane is present enough to run a focused pilot. Keep benchmark and accuracy claims out of the decision until you test against your own sources.
Run a pilot with real tenant-scoped sources, deletion checks, and recall inspection.
The link below stores the six checklist answers as query parameters. Opening that URL reproduces the same yes, partial, and no choices in any browser.
The six dimensions combine Answer Engine's vocabulary posts with public control-plane references for AI risk, MCP portability, erasure, and traceability.
NIST frames AI risk management as a practical design, development, use, and evaluation discipline.
MCP defines a standard way for LLM applications to connect with external context, tools, and workflows.
Article 17 is the regulatory anchor for right-to-erasure thinking in memory lifecycle design.
The AI Act raises expectations for traceability and logging in regulated AI contexts.
Once the missing controls are clear, compare memory systems on the same columns instead of brand language or unverified performance claims.
A score above 4.5 out of 6 is ready for a focused pilot, not a universal benchmark. The goal is to prove inspectability, supersession, forgetting, portability, multi-source recall, and tenant isolation on your own sources.
No. It is a readiness rubric for controls and buyer evidence. Do not treat it as an accuracy benchmark or publish a performance number from it.
The page encodes only the six yes, partial, or no answers in the query string so a shared link can reproduce the same result without storing user data server-side.
No. Equal weighting makes the public rubric easy to share. Regulated, multi-tenant, or customer-facing agents should usually weight deletion proof and tenant isolation more heavily in their internal review.