Agent runtime
Memory baked into a self-contained agent runtime — the one architecture this catalog explains but does not list.
Agent runtimes — Letta (MemGPT), Hermes, and the like — build memory into a complete, stateful agent as a first-class subsystem: the runtime decides when to store, what to keep, and when to inject context, alongside skills, sessions, and user profiles. The catch, and the reason this family carries no cards, is that you can't adopt the memory on its own. It only works inside that agent. To get it you run your whole agent inside the runtime — you can't bring it to Claude Code, Cursor, a LangChain app, or Claude Desktop. By this catalog's inclusion test (a memory layer must be usable independently in other agent frameworks, harnesses, IDEs, or chat interfaces), agent runtimes fall out of scope. The family stays here to name the pattern and mark the boundary.
What makes this family unique
The agent owns its memory — which is exactly why it isn't an adoptable memory layer. In every catalogued family, memory is a component you can wire into an agent, harness, IDE, or chat client you already run. In an agent runtime, memory is inseparable from the agent's own cognition, so adopting it means adopting the entire runtime. That makes runtimes a genuine and important architecture, but a different product category from the portable memory layers this site catalogs. Letta and Hermes are the canonical examples of where the line falls; the About page states the full eligibility rule.
Frameworks in this family
0 catalogued — none yet.
No frameworks catalogued in this family yet.
Use cases this family is built for
Top-down recommendations from the use-case playbook. Each names the one binding constraint that picks the tool, the primary pick (which may sit in another family when the case spans more than one), and runner-ups.
No use cases tagged for this family yet.
Last verified 2026-06-29 · updated by manual