Episodes & Historical Backfill

Episodic Memory

While the Memory Ledger tracks atomic facts about you ("Alex prefers Postgres"), Episodes track temporal context about what you actually did ("Last Tuesday, we analyzed three cloud database providers and selected Supabase for Project Phoenix").

Episodic memory enables your agents to maintain long-term continuity across weeks of work without requiring you to manually re-explain past conversations.

armes.ai

Expand

How Episode Distillation Works

When a chat thread has been inactive for a period of time:

1. Lightweight Extraction: A background summarization pass analyzes the transcript under strict Zero Data Retention inference.
2. Three-Part Episode Format:
   • Key Decisions: Final choices made during the discussion.
   • Open Threads: Unresolved questions or pending tasks.
   • Artifacts Produced: Notes created, files referenced, or code snippets generated.
3. Storage: The summary is saved as an Episode linked to that conversation ID.

Referencing Past Episodes

When you start a new conversation with any agent, the agent can reference past episodes whenever relevant:

User: "What did we decide regarding database pricing when we were comparing vendors last week?"

Agent: "In our conversation on August 28 ('Cloud Vendor Comparison'), we concluded that usage-based compute pricing was 40% more cost-effective for our forecasted query volume. You decided to begin with the Pro tier."

Historical Backfill (Opt-In)

If you imported chats from another platform or have existing conversations from before enabling memory:

1. Open Settings → Personalization → Episodes.
2. Click Run Historical Backfill.
3. The system processes threads under Zero Data Retention inference and generates structured episodes.
4. Safety Invariant: All candidate atomic facts extracted during backfill land in the Needs Review queue in your Memory Ledger. Backfilled facts never auto-activate into prompts without your explicit approval.