Reference/Agent patterns
Memory, traces, and scratchpads
Transcript vs side trace vs long-term memory. What to store, hash, and drop.
Three stores.
| Store | Lifetime | Goes to the model? |
|---|---|---|
| Transcript | this run | yes (trimmed) |
| Side trace | this run | no (logs, billing) |
| Long-term memory | across runs | only via retrieval |
Transcript
List of events: user, assistant/tool, system parse errors. This is short-term memory.
Side trace
Hashes of files, token counts, approval digests, raw HTTP (redacted). Bigger than the prompt. Required for evals.
Long-term
If you "remember the user likes Oslo", write a note and retrieve it. Do not grow a system prompt of lore. Use the RAG project.
Scratchpad
A notes string the model may rewrite each turn. Cap length. It is still untrusted text.
Hashes
sha256(content)[:12] in traces instead of full source. Compare round 1 vs round 2 without logging secrets.
Drop first
Duplicate observations, thoughts, old 429s. Keep the goal, last errors, last successful tool result.
Tip:If you cannot replay a run from the side trace, you cannot improve a run.