What is Summary Memory?
Summary memory is a memory type that stores compressed summaries of information rather than complete verbatim content, enabling agents to maintain awareness of extensive histories without consuming excessive context window space. Instead of keeping every detail of past interactions or observations, summary memory distills the essential information, key points, and important facts into compact representations that preserve the critical content while discarding less important details.
The summarization process typically uses an LLM to read through content and generate condensed versions that capture the main points. This can happen at ingestion time (summarizing documents before storing), periodically during operation (summarizing old conversation history), or on-demand when retrieving information. The summaries may be multi-level (summaries of summaries for very long histories) or selective (summarizing some content while keeping recent or important items in full fidelity).
Summary memory provides a practical solution for agents operating over extended periods or dealing with large volumes of information. It enables maintaining awareness of context that would otherwise exceed token limits, trading perfect recall for efficiency. The main challenge is determining what to preserve in summaries versus what to discard, as overly aggressive summarization loses important nuances while conservative summarization doesn't achieve sufficient compression. Many systems combine summary memory with other memory types, using summaries for distant history while maintaining detailed recent memory.