Session memory storage

Zep holds the context an AI system needs to remember across conversations. A Beam agent writes each new exchange into Zep as sessions run, reading the messages, applying your rules about what is worth keeping, and saving the relevant pieces to the right user or session record. Sensitive or low-value content is dropped per your policy. When something looks like it should not be stored, personal data outside your retention rule, or a session that cannot be tied to a known user, the agent sets it aside for a person to decide rather than persisting it by default.

Fact retrieval

When a task needs prior context, a Beam agent queries Zep for the facts tied to that user or session, reads what comes back, and supplies the pieces your rule marks as relevant to the step at hand. Stale or conflicting entries are noted instead of used blindly. If the memory returns nothing, contradicts itself, or surfaces a fact that clashes with newer data, the agent flags the gap to a person rather than acting on shaky context. That keeps downstream decisions grounded in memory your team can actually trust and defend.

Memory summarization

Over long histories, raw context grows unwieldy, so Zep condenses it. A Beam agent starts a summary when a session passes the length your team sets, reads the accumulated messages, and stores a shorter record that keeps the key facts under your rules. Older detail is archived, not lost. Summaries that would drop something your policy marks as important, or histories with conflicting facts the agent cannot safely merge, are handed to a person to check. That way the memory stays compact without quietly discarding detail a later task might depend on.