The four design principles

SSGM is built on four principles. Each one directly counters a failure from Chapter 4.

Principle 1: Pre-consolidation validation

Before any new memory is committed, it must pass through a Write Validation Gate. This gate acts as a consistency checker: it retrieves the agent's established core facts and checks whether the proposed new memory logically contradicts them. If the new memory conflicts with protected core facts, the update is rejected. The paper describes using formal logical and natural-language-inference checks for this. The purpose is to stop hallucinated or inconsistent statements from being written into long-term memory in the first place, heading off the hallucination and self-reinforcing-error failures.

Principle 2: Temporal and provenance grounding

When reading, candidate memories pass through a Read Filtering Gate that judges them on two axes. First, provenance: was this memory created by a trusted source, or could it be an injected adversarial instruction? Second, freshness: how old and how relevant is it? For freshness the paper uses a cognitive decay function (a Weibull curve) that lowers a memory's relevance score as time passes since it was last useful. Memories that fall below a freshness threshold are pruned before they reach the agent's context window. This directly attacks staleness and poisoning: the agent reasons only over fresh, trusted data.

Principle 3: Access-scoped retrieval

Retrieval must not rely on semantic similarity alone. The Read Filtering Gate also enforces identity-based access control (the paper mentions attribute-based access control) directly in the query, so an agent or user can only retrieve memories they are permitted to see. This counters the privacy-leakage failure from Chapter 4, and specifically the finding that fully connected memory networks maximize leakage. The fix is strict isolation of memory sub-graphs so that one user's or persona's data cannot bleed into another's.

Principle 4: Reversible reconciliation

To bound drift over the long run, the storage itself should be dual-track. SSGM pairs a fast, mutable "active" memory (used for everyday reasoning) with an append-only, immutable "episodic ledger" that records raw observations and acts as the source of truth. Periodically, the system replays the active memory against this ledger and corrects any concepts that have drifted. This is the formal version of the "keep raw records linked to summaries" advice from Chapters 3 and 4, and it provides a rollback mechanism if the agent's behavior degrades badly.

The whole architecture, then, is a pipeline: a write-validation gate guards what goes in, a read-filtering gate guards what comes out (by freshness, provenance, and access), and a background reconciliation process periodically cleans the mutable memory against an immutable ledger.