Memory Extraction
Memory extraction is the step in a memory pipeline where raw context — a conversation, a document, a record — is processed, usually by an LLM, into a smaller set of discrete facts or memory items to store, discarding whatever the extraction step doesn't select as salient.
Extraction is what makes a memory system scale past raw storage: instead of keeping every conversation verbatim and searchable forever, the pipeline reads new content once, pulls out what it judges worth keeping, and stores the compressed result. Mem0's published architecture is a concrete, documented example — an extraction phase combines a running conversation summary with a window of recent messages, calls an LLM to produce a set of candidate memory facts, and a separate update phase reconciles those candidates against existing memories with an add, update, delete, or no-op decision per candidate (Chhikara et al., 2025).
"Extraction loss" is simply the set of things that didn't make it through that filter. It's not a bug — it's the direct, unavoidable cost of choosing to compress at all — but it is specific and diagnosable: an extraction step tuned to pull out factual claims may readily preserve "the client took a 401(k) loan" while dropping "because of a documented liquidity concern," because the reasoning behind a decision is exactly the kind of secondary detail a salience filter tends to rank lower than the headline fact.
Extraction quality is also not purely a property of the underlying model — the same model can extract very different amounts of structure depending on the pipeline it's embedded in, since the extraction prompt, the context window it's given, and what counts as "salient" in that pipeline's design all shape the outcome independent of which LLM is doing the extracting.
When evaluating a memory system, extraction loss is measurable directly: run the system and a verbatim-storage baseline over the same corpus, and the exact-match gap per task type shows what the extraction step is dropping — a wide gap on rationale-lookup or rule-attribution tasks specifically indicates the extraction pipeline isn't preserving reasoning or citations, even if it preserves headline outcomes well.
Common pitfalls
- Assuming extraction loss is fixable purely by swapping in a stronger model — pipeline design (what the extraction prompt asks for, what context it's given) shapes what survives independent of model capability.
- Judging extraction quality from a single aggregate score rather than per-task-type — a pipeline can extract facts well and rationale poorly, and an aggregate hides that split.