Anthropic's 13-page agent memory guide lays out 5-layer architecture cutting token cost 90%
blaizedsouza · x · 2026-09-15
Anthropic released a 13-page PDF on agent memory describing a five-layer architecture:
- Working memory: the context window itself; most agents stop here and break when it fills up.
- Episodic memory: timestamped interaction logs letting the agent recall past events (e.g., why a deploy failed).
- Semantic memory: facts and entity relationships stored as a knowledge graph, persisting across sessions.
- Procedural memory: turning approaches that worked into reusable skills, skipping straight to what worked next time.
- Forgetting: deciding what to discard to control cost.
The guide claims this cuts token cost by 90% and enables agents that actually learn. The post also cites an article arguing the key skill of 2026 is "harness engineering" — the bottleneck has shifted from prompts and context to everything around the model.
More from coding & agent
- The detective novel thought experiment: why hidden states beat tokens for model handoff — CShorten30 · 2026-09-15
- oh-my-pi open-sources agent context compaction design docs, now at 31k GitHub stars — nirmal_dist · 2026-09-15
- RAFT v3.1 open-sources retrieval-augmented fine-tuning to clone human personas — jessi_cata · 2026-09-15
- 18 sub-agents cut to 1: Polylane argues sub-agents are just wrong — zeeg · 2026-09-15
- deepseek-harness: an open-source plugin-based coding agent framework, unverified — goyalshaliniuk · 2026-09-15
- DeepSeek's full-stack code output lags Claude Code and Cursor — devs trade prompting tips — AdEnvironmental4143 · 2026-09-15