Memory Schema

Pass a memory_schema to the context to type-check memory writes. When a step writes to a declared (namespace, key) pair with the wrong type, CARL raises a MemorySchemaError instead of silently storing bad data.

from mmar_carl import ReasoningContext

context = ReasoningContext(
    outer_context=data,
    api=client,
    memory_schema={
        "results": {
            "score": float,
            "label": str,
            "tags": (list, tuple),   # a tuple of types = a union
        },
    },
)

How it works

• The schema is {namespace: {key: type_spec}}.
• On every memory_write / memory_append to a declared pair, CARL runs validate_memory_write and raises MemorySchemaError (a TypeError subclass) on a mismatch.
• Pairs not in the schema pass through unchecked — you only validate what you declare.
• type_spec can be a single type or a tuple of types (a union).

This pairs well with the pre-execution check that warns when a step reads a $memory.* key no prior step writes — together they catch most memory wiring bugs before they cost you an LLM call.

See also

• Memory overview
• Memory steps