Conversation history
Conversation history is the running record of messages exchanged between the user and the agent in the current session. Distinct from long-term memory (which persists across sessions), conversation history is the working context the model reasons over for the current turn.
Why it matters
Without conversation history, every turn is independent: the user has to re-explain context every message. With it, the agent understands "what about tomorrow?" because it remembers "we just talked about Friday."
The history is technically distinct from long-term memory:
- Conversation history is the recent message exchange, kept in working memory for the current session.
- Long-term memory is the durable, consolidated record across sessions, written nightly through dreaming.
Conversation history accumulates fast. A 30-message exchange might be 10K tokens. Most modern frontier models have 100K-200K token windows, so a single conversation rarely exceeds the limit. But long sessions need archiving: older messages get pushed to long-term memory; a summary stays in the active context.
The hard parts:
- What to keep verbatim, what to summarize. Recent messages are kept as-is. Older messages get summarized into a "we discussed X, decided Y" form to save tokens.
- Cross-channel. The user might switch from SMS to voice; the conversation history should follow them.
- Privacy. Conversation history is sensitive data; it is encrypted at rest like all other user data.
How Ajeris handles this
Ajeris's agent loop maintains conversation history as part of working memory, archived into long-term memory when the active session ends or context approaches the model's window. The history is per-user and follows the user across channels: starting on SMS and continuing on voice keeps the same history available.
Older content gets summarized rather than dropped, so the agent never abruptly forgets a long conversation. The summarized version has lower fidelity than the verbatim original, which is the trade-off.
