Skip to main content
Memory lets the agent maintain context beyond the current message — either within a single conversation or across multiple sessions over time.

Memory strategies

None (default)

No memory. Each message is treated as an independent, stateless interaction. The agent has no awareness of previous turns. Best for: Single-turn tasks, lookup agents, classification agents, any use case where conversation history doesn’t matter.

Short-term memory

The agent remembers the current conversation session. Previous messages in the same session are injected into the prompt context.
ParameterDescription
Max TokensMaximum token budget for injected history. Older messages are dropped when the budget is exceeded.
Metadata FilterOptional filter to scope memory retrieval by session metadata (e.g., only recall messages tagged with a specific topic)
Best for: Customer support conversations, multi-turn Q&A, any interaction where the agent needs to remember what was said earlier in the same session.

Long-term memory

The agent persists context across sessions using one of two runtime strategies — Neuralyzer or CoD Summarizer (cod-summarizer). These are summarization / compression pipelines managed by the platform. Choose Memory Strategy:
StrategyWhat it does
NeuralyzerContinuously summarizes and compresses conversation history so older turns collapse into durable context instead of growing without bound.
CoD SummarizerChain-of-Density summarization: builds compact summaries and can store them in a knowledge-base collection you name (see Collection ID below).
Shared parameters (both strategies):
ParameterDescription
Interaction Threshold (Qty)Number of interactions before a summarize step runs. Use -1 for no limit (always follow runtime defaults).
Interaction Threshold (Tokens)Token budget threshold before summarizing. Use -1 for unlimited.
Interaction KeepHow many recent interactions to leave unsummarized. 0 means everything eligible can roll into summarized form.
CoD Summarizer only:
ParameterDescription
Collection IDKnowledge-base collection where summaries are written (must match how your runtime/knowledge stack expects storage).
CoD Max LoopsMaximum refinement loops for the Chain-of-Density pass (default in UI: 5).
Best for: Assistants that should remember a user or account across returning sessions, without relying on raw transcript replay alone.
Long-term behavior is strategy-driven (Neuralyzer vs. CoD Summarizer), not “pick Top K / similarity on an embeddings index” in Studio. If you need retrieval over uploaded documents, that is Knowledge Base + Embeddings.

Combining memory with knowledge base

Short- or long-term memory and Knowledge Base (RAG over documents) can be active together: the agent can combine conversation-side context with document embeddings in the same reply. CoD Summarizer in particular may write summaries into a collection that participates in your knowledge stack; that is separate from configuring an embeddings model for long-term memory in the Memory UI.

Next steps

Dev Mode

Access advanced controls for the system prompt and agent configuration.