Core terms and abbreviations recurring across the two papers
The following terms are organized alphabetically, covering both Mem0 and ACE research lines.
General
| Term | Full Name / Meaning |
|---|---|
| LLM | Large Language Model |
| RAG | Retrieval-Augmented Generation |
| ICL | In-Context Learning (few-shot or many-shot) |
| KV cache | Key-Value cache, critical caching mechanism for long-context inference |
| ReAct | Reason + Act, an agent paradigm alternating reasoning and action |
Mem0 Related
| Term | Meaning |
|---|---|
| Mem0 | Basic memory architecture storing facts in natural language |
| Mem0g | Graph memory variant of Mem0; nodes are entities, edges are relations |
| Extraction phase | Extracts candidate facts from message pairs |
| Update phase | Evaluates candidate facts and executes ADD/UPDATE/DELETE/NOOP |
| LOCOMO | Long-term conversational memory evaluation dataset |
| LLM-as-a-Judge (J) | Uses another LLM to score as an evaluation metric |
| F1 / BLEU-1 | Traditional lexical overlap metrics |
ACE Related
| Term | Meaning |
|---|---|
| ACE | Agentic Context Engineering |
| Playbook | Evolving context that accumulates reusable strategies and domain insights |
| Generator | Produces reasoning traces for new problems |
| Reflector | Distills reusable experiences from traces |
| Curator | Merges experiences into structured delta updates |
| Delta update | Incremental update that only modifies relevant bullets, avoiding full rewrites |
| Grow-and-refine | Expand-first, then deduplicate context maintenance mechanism |
| Brevity bias | Over-compression leading to loss of domain details |
| Context collapse | Full rewrite causing drastic information loss |
| AppWorld | Interactive agent evaluation benchmark |
| FiNER / Formula | XBRL financial domain evaluation benchmarks |
Evaluation Metrics
| Metric | Description |
|---|---|
| TGC | Task Goal Completion rate (AppWorld) |
| SGC | Scenario Goal Completion rate (AppWorld) |
| p50 / p95 | Median / 95th percentile latency |
| Rollouts | Number of invocations during optimization |