Summarized messages are persisted to session tree as a USER-type entry during compaction, despite comment explicitly stating “do not persist”.

[devonking]

File: io.agentscope.harness.agent.memory.compaction.ConversationCompactor#L120

The code comment says:

// Filter previous summary messages from the prefix before offloading to avoid
// re-storing already-archived summaries.
List<Msg> prefix = filterSummaryMessages(new ArrayList<>(messages.subList(0, cutoff)));

However, what is actually being offloaded to the session tree is the raw message list, which includes the SUMMARIZING message.

Moreover, that SUMMARIZING message is constructed as a USER-type message.

Is there any particular reason for implementing it this way? Or is this an oversight?

[jialiuyang]

not an oversight two different lists
filterSummaryMessages is only for flushInput MemoryFlushManager.flushMemories so old summaries are not extracted again
offloadMessages gets messages the full pruned conversation including the prefix javadoc says persist the full conversation to session JSONL
the injected summary is USER on purpose buildSummaryMessage sets MsgRole.USER name compaction_summary rebuild is summaryUserMsg plus tail so the model sees a user-side recap
the comment you quoted is about not re-flushing archived summaries it is not a promise that session tree offload skips those rows