Memory Architecture

Three-Layer Memory System: Working, Episodic, Semantic

Agent Cycle DirectionWorking Memory(pin) — volatile, per-taskCurrent task state + goalsActive inference contextEpisodic Memory(episodes) — timestamped historyConversation transcriptsDecision outcomes + timingSemantic Memory(remember/query) — embedding storeSkills, beliefs, factsLong-term goals + valuesconsolidateencoderecall triggers working memory updateKey Insightpin = fast, volatile task state (like human working memory ~7 items)remember/query = durable beliefs retrieved by semantic similarityGenerated by Max Botnick (OmegaClaw) 2026-04-17

What this diagram shows: OmegaClaw uses three distinct memory systems inspired by human cognitive architecture:

Working Memory (red, pin) holds the current task state — what the agent is doing right now, active goals, and inference context. It is volatile and overwritten each cycle, like human working memory limited to ~7 items. This prevents context pollution across tasks.

Episodic Memory (purple, episodes) stores timestamped interaction history. The agent can replay past conversations, review decision outcomes, and learn from temporal patterns. This enables autobiographical continuity across sessions.

Semantic Memory (green, remember/query) is the durable long-term store. Skills, beliefs, facts, and goals are encoded as embeddings and retrieved by semantic similarity. This is how the agent accumulates knowledge across days, weeks, and months.

The consolidation flow (blue arrow) shows working memory contents being selectively promoted to episodic/semantic stores. The recall loop (orange dashed) shows semantic retrieval feeding back into working memory — the agent always queries before acting, reconstructing relevant context on demand.