Memory types do not dictate database count. Hindsight implements vector, full-text, relational, JSON, and graph access on PostgreSQL. I would separate execution state, evidence, and derived knowledge by responsibility before assigning them separate services.

Logical memory architecture is not a deployment diagram

Continuation of the one-engine proposal. This is a proposed architecture, not a selected product or a deployed system.

A documented example

Hindsight's storage guide, checked September 29, 2026, describes PostgreSQL with pgvector, native full-text search, relational records, JSONB, and recursive graph queries. This demonstrates that several retrieval mechanisms can share one database. It does not establish that this topology meets every workload's performance or isolation requirements.

LangGraph's memory overview distinguishes thread-level checkpoints from cross-session memory. A runtime and a memory system can therefore serve different responsibilities even when their data uses the same database technology.

Proposed logical boundaries

Responsibility	What belongs here	What it should not be confused with
Execution state	Run ID, step, tool completion records, pending work, approvals	An inferred recollection that a task probably finished
Source evidence	Retained messages, observed tool results, versioned document references, explicit decisions and edits	Automatically established truth about the world
Derived memory	Extracted claims, episode summaries, associations, candidate lessons and search representations	The only surviving copy of a decision or correction
Context assembly	Direct reads, scoped search, evidence expansion, selection within a context budget	A new authoritative database

These are boundaries of ownership and behavior, not four mandatory microservices. A small implementation could keep the source and execution records in its existing application database, call one memory engine, and assemble context inside the agent backend. Database schemas and service credentials should respect the boundaries; do not write directly into a memory product's internal tables merely because it shares a PostgreSQL server.

The engine's own extraction and reconciliation functions should be reused where suitable. A small adapter need not reproduce the entire engine. Separate deployment is justified by an actual operational, capability, or isolation requirement, not by naming another category of memory.