CHAT WINDOW
Memories: Built for Ongoing Relationships
Long-term memory for Tavus PALs, built around the relationship between a PAL and the person talking to it, so every conversation builds on the last.
Memories gives Tavus PALs long-term memory, built around the relationship between a PAL and the person talking to it. Each relationship has its own evolving memory of who that person is, what matters to them, and what has happened across their conversations. That understanding carries forward when they return, so they can pick up where they left off instead of explaining themselves again.
Finding an old fact is useful, but a PAL also needs to know when that fact should change the conversation: which question to ask, which explanation to use, or which suggestion makes sense for this particular person. Mention an upcoming board meeting in passing, and when you return a few days later your PAL asks, “How did the board meeting go?” You never asked it to remember. It connected something you shared before with the conversation happening now.
For the thinking behind it, read the full announcement.
Three things define Memories
- Built around the relationship, not just facts. Every PAL-person relationship has its own Memory Store with two parts: a Profile of the person (preferences, goals, communication style, longer-term context) and a Timeline of what has happened (conversations, decisions, commitments, and changes over time). Both are handed to the PAL the moment a conversation begins.
- Selective memory, shaped by the PAL’s role. A tutor and a coach shouldn’t build identical memories of the same person. Memory is maintained through consolidation, not accumulation: after each conversation the PAL revises its understanding rather than stacking up summaries. Recent conversations stay detailed, older ones are progressively consolidated, and the PAL can search saved history when an exact detail matters.
- Automatic memory, with control over what carries forward. The Profile and Timeline are inspectable, correctable, and deletable. Developers can also supply Pinned Memories: explicit context that takes priority over learned memory and can’t be rewritten, so a tutor can start with notes from a teacher, or a support PAL with a customer handoff.
The result is a shift from a tool you have to keep directing to a counterpart you can work with. A goal can span multiple conversations. A lesson can build on a previous breakthrough. A project can move forward without starting over.
What this unlocks
- A K-12 tutor brings back the pizza-slice explanation that helped a student understand fractions when it introduces percentages, and knows which parts still need practice.
- A patient-education PAL follows someone from before surgery to after it, asking whether their sister was able to stay with them instead of repeating pre-op questions.
- A support PAL remembers the customer doesn’t have a screwdriver and starts with the checks that don’t require opening the device.
- A coach follows up on the obstacle, not just the goal: “Were you able to keep Tuesday morning free, or did something get scheduled over it?”
- A sales or customer-success PAL picks up at the actual point of decision, asking whether the data-retention questions blocking security approval were resolved instead of restarting the pitch.
How it works
Connect conversations with a consistent participant_tag. Tavus combines your account, the PAL, and the tag to resolve the right Memory Store, keeping every relationship separate. After a conversation ends, the Profile and Timeline are updated asynchronously. When the next conversation starts, they’re assembled into context alongside any Pinned Memories, and a retrieval agent can search saved conversation history when more detail is needed. Tavus maintains the infrastructure, so you don’t have to build memory from scratch.
Built for conversation, tested on recall
Memories reached 93.4% on LongMemEval-S, a benchmark for long-term conversational memory spanning factual recall, preferences, knowledge updates, temporal reasoning, and reasoning across conversations: 500 questions with 38 to 62 historical sessions behind each one, 23,867 sessions in total. See the full results.
Beyond a single conversation
Give your PAL a memory of the people it talks to. Try Memories with your PAL, then come back and pick up where you left off.
Try Memories
Learn how to build:
Every conversation should build on the last.
Try Memories with your PAL, then come back and pick up where you left off.
Frequently Asked Questions
Memories gives Tavus PALs long-term memory built around each PAL-person relationship. Every relationship has its own Memory Store with two parts: a Profile of who the person is (preferences, goals, communication style, and longer-term context) and a Timeline of what has happened across their conversations (events, decisions, commitments, and changes over time).
Developers connect conversations with a consistent participant_tag. After a conversation ends, the Profile and Timeline are updated asynchronously. When the next conversation begins, they’re loaded into context alongside any Pinned Memories, and a retrieval agent can search saved conversation history when an exact detail matters.
The Knowledge Base uses retrieval-augmented generation (RAG) to find relevant information in a body of documentation, which is exactly the right tool for that problem. Memories solves a different one: understanding how a person’s history should influence a live conversation.
Rather than running a search step before every response, Memories keeps a compact, maintained understanding of the person and your shared history ready at the start of each conversation, and only looks back through saved history when more detail is needed. Use the Knowledge Base for what your PAL should know. Use Memories for who it’s talking to.
Yes. Memories is built for continuity across conversations. Recent conversations stay detailed while older interactions are progressively consolidated into broader summaries, so the PAL understands how things have developed without carrying every sentence. Changes keep their history too: when someone changes jobs, the PAL understands that one role was true, something changed, and another is now current. When an exact detail matters, the PAL can look back through saved conversation history.
Yes. What a PAL learns is inspectable, correctable, and deletable. You can review the Profile and Timeline, fix a misunderstanding, or remove information that should no longer be part of the PAL’s memory.
Developers can also supply Pinned Memories: explicit context that takes priority over learned memory and can’t be rewritten by the learning process. A tutor can start with notes from a teacher. A support PAL can start with context from a customer handoff.
You can keep using it. If you want your PAL to benefit from what you already know about a person, pass that context in as Pinned Memories, which take priority over learned memory and are never overwritten by the learning process. Tavus maintains the underlying memory infrastructure, so you don’t have to build consolidation, retrieval, or long-term storage from scratch.
We evaluated Memories on LongMemEval-S, a benchmark for long-term conversational memory covering factual recall, preferences, knowledge updates, temporal reasoning, and reasoning across conversations. Across 500 questions with 38 to 62 historical sessions behind each one (23,867 sessions in total), Memories reached 93.4%.
For context, published results from dedicated memory providers include ByteRover at 92.8% on LongMemEval-S and Mem0 at 94.4% on LongMemEval. These are contextual comparisons using different models and evaluation procedures, not controlled head-to-head tests. See the full results.