docs/deterministic_memory_stack_v0.md

Deterministic Memory Stack v0

Clean product one-liner: the repo now has a typed memory runtime over its fast exact lookup and structural discovery floors, but it still needs front-door routing and dedicated benchmarks to make that memory layer a fully governed product surface.

Current Read

Layer	Current state	Why
exact lookup	present	the hot graph already supports very fast exact lookup
structural discovery	present	hologram retrieval works as the fallback structure search layer
auditability	present	receipts and deterministic kernel reads already exist
typed ledger	present	law, schema, and compiled runtime now exist
unified memory product	partial	recall is unified in the shell, but not yet routed through the conversational front door

Design Rule

Use this order:

exact lookup -> temporal lookup -> relational expand -> structural discovery

Not:

semantic retrieval first

Why in Plain English

If the system already knows the exact thing, it should fetch that exact thing.

If the question is about the past, it should retrieve the right version from the right time.

If the answer depends on rationale or provenance, it should follow the recorded relations.

Only when those fail should it widen into probabilistic discovery.

Shipped Surfaces

• policy/typed_memory_ledger_v0.json
• schemas/typed_memory_unit_v0.json
• benchmarks/memory_architecture_scorecard_v0.json
• build/system/typed_memory_runtime.json
• build/system/memory_readiness.json

Bottom Line

You already have the fast lookup floor and now a real typed runtime over it.

The next product move is to route exact and temporal memory questions through that runtime by default and put it on its own benchmark gate.