← All posts

Agents that remember you across conversations

Long-term memory lands: agents extract facts and preferences from past conversations and recall them in brand-new ones, with continuity that survives sessions.

Vocion Teamvocion-v1.74.0

Tell an agent once that your emails must stay under 75 words, and every future conversation should already know it. With v1.73 and v1.74, agents on the BYOA runtime get exactly that: two layers of memory backed by AWS AgentCore Memory, both opt-in, both built to never break chat.

This is different from the mission working notes that shipped in v1.60. Those are a team's memory between scheduled checks. This is the chat loop's memory of you: what you said, what you prefer, carried across sessions.

What's new

Why we built it

Stateless chat loops rediscover the user every session. Anything you told the agent yesterday (constraints, formats, context about your accounts) has to ride along in the conversation payload or be repeated. Continuity fixes the mechanical half: the loop keys a Memory session per persisted conversation and owns its own history. Long-term memory fixes the human half: extraction strategies run against completed turns and produce records that survive the conversation that created them.

The default continuity mode is belt and suspenders: payload history still rides along, and the loop prefers whichever source is at least as complete. That protects conversations that predate Memory and survives failed writes. Set VOCION_MEMORY_AUTHORITATIVE=1 to omit payload history and take the real token savings.

How to use it

Provision the Memory store once (idempotent; it also creates the two extraction strategies):

ENV=dev bash infra/agentcore/provision.sh

Then enable it on the runtime artifact with a single env var:

VOCION_AGENTCORE_MEMORY_ID=<memory-id>

That's it. From here the behavior we verified live against the real SSE route: state a preference in one conversation ("emails strictly under 75 words"), and the extraction strategy turns it into a structured record in about a minute. Open a brand-new conversation and the agent recalls it correctly. The recall crossed sessions through long-term memory, not chat history; with VOCION_MEMORY_AUTHORITATIVE=1 set and payload history omitted, the agent answered from Memory alone.

Records are namespaced per user (/facts/<actor>, /preferences/<actor>), so one person's preferences never leak into another's context.

What's next