RFC-0007 Memory

RFC-0007 — Memory

Title: Memory Extension Contract

Status: Accepted

Summary

Defines the Memory subsystem — a pluggable persistence layer for execution contexts that enables conversation history, long-term recall, and provider-swappable storage.

Motivation

The Brain must remember past interactions to provide coherent multi-turn conversations. Rather than hard-coding a storage backend, Xyberos defines a contract so deployments can choose in-memory, SQLite, Redis, or vector stores without changing the pipeline.

Contract

class Memory(ABC):

    @abstractmethod
    def retrieve(self, context: object) -> Any:
        """Return memory relevant to the supplied execution context."""

    @abstractmethod
    def store(self, context: object) -> None:
        """Persist information from the supplied execution context."""

The contract intentionally takes object rather than CognitiveContext, keeping extension contracts independent of core layer types. A provider may inspect the object as needed.

Pipeline Integration

The Brain calls retrieve before LLM generation and store after a response is produced:

Memory (retrieve) → Knowledge → Planner → Tools → LLM → Memory (store)

Providers

Provider	Backend	Persistence
InMemoryMemory	Python list	Process only
SqliteMemory	SQLite file	Cross-restart

Both implement the same contract. Swap via:

app = create_app(memory=SqliteMemory("chat.db"))

Future Providers

• Redis — shared state across processes
• Vector store — semantic/embedding-based retrieval