How to keep the last N messages verbatim when using EventsCompactionConfig?

[jimkont]

Hi all,

I’m using ADK’s EventsCompactionConfig with an LlmEventSummarizer and had a question about how to preserve recent conversation turns.

My goal is to always keep the last N user interactions (e.g., 3) fully verbatim in the session context, and only summarize the older history before them. Conceptually:

[ summarized history ... ] + [ last 3 interactions verbatim ]

From the docs and behavior I’m observing:

• compaction_interval is based on user-initiated invocations
• overlap_size creates overlap between consecutive compaction windows

but overlap_size does not seem to guarantee that the last N interactions remain un-summarized — they eventually get re-summarized in later compactions

So my questions are:

• Is there a supported way in ADK to “protect” the last N interactions from compaction, so they always remain verbatim?
• Or is compaction strictly window-based, meaning every invocation will eventually be summarized as the window slides forward?

If the latter, what is the recommended pattern to preserve recent verbatim context (e.g., storing recent turns in session state or a separate memory layer)?

My use case is an analytics agent where recent wording and exact criteria matter for follow-up questions, so keeping a small verbatim working set would be very helpful.

Thanks in advance!

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: It seems that ContextFilterPlugin is what you are looking for.

---

Hi @jimkont,

Thanks for your detailed question. Your understanding of how EventsCompactionConfig works is correct.

To directly answer your questions:

1. Is there a supported way in ADK to “protect” the last N interactions from compaction?
   It appears so. While EventsCompactionConfig itself manages the summarization window, there is a plugin called ContextFilterPlugin that seems designed for your exact use case. In a sample "hello world" app in the ADK repository, there is a line that shows how to use it: ContextFilterPlugin(num_invocations_to_keep=3) [1]. This plugin would be added to your agent to filter the context that gets sent to the LLM, ensuring the last N invocations are always included in their original, verbatim form.

2. What is the recommended pattern to preserve recent verbatim context?
   The recommended pattern appears to be using ContextFilterPlugin as described above.

We hope this helps you build your analytics agent!

[1] https://github.com/google/adk-python/blob/main/contributing/samples/hello_world_app/agent.py

[jimkont]

can you clarify how the ContextFilterPlugin and the EventsCompactionConfig be combined together?
my use case it so compact the session but not the last X (e.g. 3) messages. From what I understand these can be used separately but not together