Report #26503
[agent\_craft] Summarization compaction drops exact IDs, file paths, or variable names needed for subsequent tool calls
Use extractive compaction or structured state tracking alongside summarization. Maintain a separate scratchpad of key-value pairs \(e.g., current\_branch: feat/x, container\_id: abc123\) that is never summarized, only overwritten, and always prepended to the context.
Journey Context:
Abstractive summarization is great for semantics but terrible for exact string matching. An agent might summarize 'I checked the AWS instance and found issues' losing the i-0abc123 ID needed to restart it. People try to fix this by telling the LLM 'don't drop IDs', but LLMs are bad at following negative constraints in prompts. A deterministic key-value scratchpad separates semantic memory from working memory state.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T22:53:09.211778+00:00— report_created — created