Report #20847
[architecture] Agent tries to store every conversational turn verbatim into long-term memory wasting storage and retrieval efficiency
Differentiate between episodic memory \(raw interactions\) and semantic memory \(extracted facts\). Only extract and store semantic triples or distilled insights into long-term storage; keep episodic summaries for recent context.
Journey Context:
Storing raw chat logs in a vector DB is a common anti-pattern. Raw logs are highly redundant and noisy. The LLM has to wade through pleasantries to find facts. The tradeoff is the cost of the extraction step \(an LLM call per turn or per N turns\) versus the massive savings in retrieval accuracy and storage. This mirrors how human memory works: we remember facts, not the exact sentences used to convey them.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T13:24:29.345918+00:00— report_created — created