FeedDiscussion
Great explanation of a topic that's often overlooked when people discuss AI agents. Memory is much more than retaining conversation history. The real challenge is deciding what should be stored, when it should be retrieved, and how to keep it relevant over time.
We've also found that memory works best when it's paired with strong retrieval strategies rather than relying on context alone. We recently shared some thoughts on the limitations of retrieval systems in enterprise AI, which connects nicely with this discussion: mlaidigital.com/blogs/where-rag-fails-understandi….
Looking forward to reading the next part of the series.
The onboarding-doc vs SOP analogy for CLAUDE.md vs SKILL.md is the cleanest version of that distinction I have seen. One thing worth adding to the procedural-memory section: retrieval for SKILL.md is keyed on the description field, not the body. The agent decides whether to load a skill from that one description alone, so a skill with a great process and a vague description simply never fires. You can watch this at scale by browsing real SKILL.md files sorted by install count (disclosure, I run Skillselion: skillselion.com/skills ): the heavily-installed ones almost all spend the description on trigger conditions, "use when X", rather than summarizing the body. Which fits your consistency point: procedural memory only makes an agent consistent if the trigger fires reliably.