Long Memory — Summarization
Configure automatic memory consolidation to reduce storage and improve search quality.
The Summarization tab controls how memories are automatically consolidated. Over time, individual conversation records accumulate and create noise in search results. Summarization merges related memories into compact, high-quality summaries.
Enable Summarizer
Activates the automatic summarization pipeline. When enabled, related memories are periodically merged into consolidated summaries.
How it affects your pipeline:
- Reduced storage — Multiple related records become one summary, lowering vector database costs
- Improved search quality — Summaries provide denser, more complete context than fragmented individual messages
- Lower token consumption — Fewer, richer memories mean smaller context windows when retrieving for your agent
- Processing overhead — Summarization uses LLM calls, which adds latency and cost during consolidation
Summarization runs asynchronously. It does not block store or search operations.
Schedule Trigger
Run summarization on a fixed schedule regardless of memory count.
Schedule Options
| Frequency | Best For |
|---|---|
| Hourly | High-volume agents with rapid memory accumulation |
| Daily | Standard chatbots and support agents |
| Weekly | Low-frequency agents or knowledge accumulation use cases |
| Monthly | Archival-style memory where consolidation is occasional |
How it affects your pipeline: Scheduled summarization ensures consistent memory quality. Even if memory count is low, periodic runs catch subtle relationships between memories that threshold-based triggers might miss.
Threshold Trigger
Run summarization when the record count exceeds a specified threshold. Range: 10 to 1000 records.
How it affects your pipeline:
- Reactive consolidation — Summarization only runs when memory volume justifies it
- Cost efficiency — No LLM calls during low-activity periods
- Burst handling — After a spike in conversations, summarization kicks in automatically
Recommendations by volume:
| Monthly Memory Volume | Recommended Threshold |
|---|---|
| < 100 records | 50 — trigger early to keep things tidy |
| 100 — 1,000 records | 100 — 200 |
| 1,000 — 10,000 records | 300 — 500 |
| 10,000+ records | 500 — 1000 |
Both triggers: You can enable both schedule and threshold simultaneously. This provides robust summarization — periodic runs maintain quality while threshold catches volume spikes.
Summarizer System Prompt
Custom instructions for the AI model that performs summarization. This prompt controls how memories are merged, what information is preserved, and the tone of the resulting summaries.
How it affects your pipeline:
- Information retention — A well-crafted prompt ensures critical details survive consolidation
- Summary format — Control whether summaries are narrative, bullet-point, or structured
- Domain adaptation — Include domain-specific instructions to preserve technical accuracy
Example prompts:
# For customer support
"Summarize customer interactions, preserving: issues reported, resolutions provided, and customer sentiment. Use concise bullet points."
# For research assistants
"Consolidate research findings into a coherent knowledge summary. Preserve key facts, citations, and unresolved questions."
# For personal assistants
"Create a concise summary of user preferences and habits. Focus on actionable information the assistant can use."
Test your prompt: Run summarization manually with different prompts before deploying. Poor prompts can lose critical information during consolidation.
Next Steps
Proceed to Compliance to configure PII redaction and heat scoring.