Personal memory
Preference ownership, short retention, correction and wrong-person recall. Measure accepted advice per dollar with consent and deletion.
RELATION-AWARE AGENT MEMORY 001
See the context.
See the decision.
Trace every source, relation, policy gate and outcome in one inspectable graph. Explore how the same context plane behaves for a person, a merchant, and an enterprise response team.
01 / 03Original graph explorerTyped relations · source provenance · scope boundary · policy selection · explainable pathAll records on this page are synthetic. The page does not run a GraphRAG service or train a model.
ATLAS / INTERACTIVE EVIDENCE GRAPH
Personal decision graph
Drag a node to rearrange · drag the canvas to pan · scroll to zoom · click an edge to inspect its relation
CONSUMER / SYNTHETIC
admitted to context
candidate or process node
denied by scope
0 nodes · 0 relations
BEYOND THE PICTURE
One graph. Three distinct operating realities.
Every scenario uses a different task, objective and failure cost. The explorer makes those differences visible; the production gates remain separate.
Operational memory
Fresh inventory, explicit margin rules and human approval. Measure accepted workflow outcomes, rework and true operating cost.
Federated context
Evidence lineage, incident handoff, tenancy and recovery. Role and region ABAC, distributed state and live connectors remain production work.
LEARNING SYSTEM / TARGET ARCHITECTURE
Observe → label → evaluate → promote
The Python kernel implements scope-limited lexical/one-hop graph retrieval and a synthetic UCB1 policy selector. Supervised prediction, unsupervised drift detection, graph ML, causal experiments and sandbox RL are specified as future evaluation tracks, not active services.
Read metrics, release gates and claim limits ↗01 Provenance-linked records
02 Independent outcome labels
03 Fixed offline comparisons
04 Guardrailed experiment
05 Human promotion or rollback
PROJECTS REFERENCED / NO IMPLIED ENDORSEMENT
Architecture and interoperability references: Cognee for graph-backed memory; LangGraph and Deep Agents for agent workflows and context; A2A and MCP for protocols; vLLM for inference serving; Google AX for execution lifecycle; OpenManus for computer-use workflows; and our candidate context-graph-compact compiler. None supplied this visualization. See the exact integration status.