Browse Platform
Platform · Reference

Agent Memory V2 Node

Beta lossless agent memory. Persists complete session JSONL values before replacing oversized active-context content with compact, retrievable references.

3 min read

:::warning This node is in beta and may not be fully tested or documented. Use with caution. ::: Beta lossless agent memory. Persists complete session JSONL values before replacing oversized active-context content with compact, retrievable references.

Properties

LabelPropertyTypeDescriptionDefault ValueIs Required
Memory KeymemoryKeybotdojo/types/stringagent_memory
Max Context Window (tokens)maxContextWindowbotdojo/types/numberHard cap on the context budget in tokens. The active model window is clamped down to this value when it is larger; a smaller model window is used as-is.128000
Model Context Window %modelContextWindowPercentbotdojo/types/numberFraction of the active model context window available to memory0.9
Summarization Trigger %compressPercentagebotdojo/types/numberSummarize older conversation history at this fraction of the context budget0.85
Post-Summarization Target %compactionTargetPercentagebotdojo/types/numberSoft active-memory target after summarization; atomic recent tool groups remain intact0.3
Recent Tail % of ThresholdcompactionRecentTailPercentagebotdojo/types/numberToken budget for exact recent messages, measured as a fraction of the summarization threshold0.2
Recent Tail Soft Limit MultipliercompactionRecentTailSoftLimitMultiplierbotdojo/types/numberAllows a complete recent atomic group to extend beyond the target tail budget1.5
Minimum Recent GroupscompactionMinRecentGroupsbotdojo/types/numberMinimum number of newest message/tool groups retained exactly2
Summary Context Limit %compactionSummaryPercentagebotdojo/types/numberMaximum fraction of the memory context budget available to the generated checkpoint0.05
Summary Source RatiocompactionSummarySourcePercentagebotdojo/types/numberDesired summary-token budget as a fraction of the historical tokens being summarized0.2
Summary Minimum TokenscompactionSummaryMinTokensbotdojo/types/numberMinimum checkpoint budget before applying the context-percentage and absolute caps1000
Summary Maximum TokenscompactionSummaryMaxTokensbotdojo/types/numberAbsolute generated-checkpoint token ceiling10000
Minimum Messages to SummarizeminMessagesToCompressbotdojo/types/numberSkip summarization when fewer historical messages are eligible5
Retrieval Preview Window (chars)maxDefaultWindowbotdojo/types/numberCompatibility setting retained in serialized V2 configuration1000
Enable Token CountingenableTokenCountingbotdojo/types/booleanUse tokenizer-based counts instead of the fast character estimatefalse
Compaction Language ModellanguageModelbotdojo/types/language_modelLanguage model used only when active session history crosses the compaction thresholdnull
Compaction PromptcompactionPromptbotdojo/types/stringInstructions used to turn older session messages into a compact context checkpoint. Leave blank to use the built-in default prompt; provide text to override it.
Externalize At (chars)externalizeAtCharsbotdojo/types/numberPersist and immediately replace very large tool outputs with compact JSONL references6000
Preview (chars)previewCharsbotdojo/types/numberMaximum deterministic preview before a compact memory reference300
View Inline Limit (chars)viewInlineCharsbotdojo/types/numberKeep already-bounded workspace view chunks inline up to this safety limit15000

Output

LabelPropertyTypeDescriptionDefault Value
Memorymemorybotdojo/interface/memoryStandalone Agent Memory V2 instancenull