​

Summaries

​

In plain terms

The analogy: turning ten pages of meeting notes into a single sticky note. You lose the word-for-word detail but keep the gist.

​

When to use it

• Long-running conversations — support cases, ongoing projects, multi-day threads. The agent remembers the arc of the discussion without carrying every message.
• Cost control on chatty sessions — instead of re-sending 200 messages each turn, the agent sends a compact recap plus the recent messages.

​

When to turn it off

• Strictly short conversations that never exceed maxMessages — there’s nothing to summarize, so it’s harmless either way, but you can disable it to skip the (rare) summarization LLM call:

memory: { storage, summaries: false }

• When you need verbatim history for audit/legal reasons — but note: summaries don’t replace the raw transcript, they supplement it. The full session is still stored; summaries are an additional compact layer.

​

Configuration

// Default (ON)
memory: { storage }

// Fewer, shorter summaries — leaner prompts, lower cost
memory: { storage, summaries: { maxCount: 5, maxTokens: 1000 } }

// Disable entirely
memory: { storage, summaries: false }

• maxCount — each time the session overflows maxMessages, one summary is created. More summaries = the agent remembers more of the conversation’s history, but old ones eventually drop. Raise it for threads that span days; lower it for short chats.
• maxTokens — directly affects cost, since these tokens are sent on every run. Lower it (1000) to save money; raise it (4000) when the conversation is rich and the agent needs more of its history to respond well.

​

How it works

1. A conversation grows past maxMessages (see Sessions).
2. The oldest messages are handed to a model (use a cheap one via memory.model) which writes a short recap.
3. The recap is stored as a summary; the original messages are removed from the active thread.
4. On future runs, summaries are injected newest-first within the maxTokens budget — so the most recent context always survives if space is tight.

​

Cross-references

• Sessions & History — where overflow comes from
• Context Budget — how summaries compete with other memory for prompt space
• Memory Overview — the full memory lifecycle