ACT-R memory equation vs. simple recency decay: crossover comes at 2.7 hours, not much gain
memokris · reddit · 2026-10-12
Comparing the last-access decay used by Generative Agents (0.995/hour) and LangChain's time-weighted retriever (0.99/hour) against ACT-R's base-level memory equation, which sums power-law traces of every past use. Key findings:
- With default d=0.5, a one-off mention leads a repeated habit for about 2.7 hours before being overtaken; with Twitter-fitted d=1.25–1.7, crossover shifts to 2–3.7 days.
- The equation traces to Anderson & Schooler (1991), who studied word reappearance in NYT headlines, child-directed speech, and inbox senders.
- Simpler rules do surprisingly well: ranking hashtags by recency comes within a few percent of ACT-R; as a cache eviction policy, the Anderson-Schooler model barely beat random while LFU got closer to optimal. Contextual inputs mattered more than the decay curve.
The author proposes drawing a log-log plot of topic recurrence vs. days since last mention from agent logs — something not yet published — before picking a decay for agent memory.
More from coding & agent
- Dev builds AI-written tool to visualize agent traces in Perfetto and find slow tool calls — DanielLockyer · 2026-10-12
- Stanford CS329Z releases Tool Use and Frameworks & Orchestration agent engineering lectures — stanfordnlp · 2026-10-12
- spaCy creator Ines Montani launches Ellf: a hybrid architecture that keeps your data on your own compute — _inesmontani · 2026-10-12
- Vibe-coded in 4 hours: a step-by-step physical-task simulator built with Claude Code — burhop · 2026-10-12
- Qwen releases QwenGyre: elastic GPU scheduling speeds long-horizon agent RL training by up to 1.85x — Alibaba_Qwen · 2026-10-12
- Hugging Face's transformers serve can turn 250 LLMs & VLMs into logit-based decision models — LysandreJik · 2026-10-12