Vector-based memory retrieval and graph-based memory retrieval aren't just different implementations of the same idea — they fail on genuinely different kinds of queries, which argues for combining them rather than picking one.
Multiple agent instances writing to the same user's memory store concurrently can produce contradictory or lost writes the same way any concurrent system can — and the fix needs to account for the fact that the "writer" reasoning about a conflict is sometimes another LLM call.
Most agent memory implementations store everything as one undifferentiated stream of facts. Separating episodic from semantic memory, the way cognitive science distinguishes them, changes both what gets retrieved and how well it ages.
MCP went from a niche protocol to 97 million monthly SDK downloads in 16 months. Here's how it actually works, what enterprise adoption looks like today, and why so many pilots still aren't reaching production.
An agent memory system that never forgets anything isn't more capable — it's a growing liability of stale, contradictory, and increasingly irrelevant information competing for retrieval against what actually matters now.
Once an organization has more than a handful of MCP servers, "which servers exist, who owns them, are they vetted" becomes unanswerable without a registry — the same problem an internal platform team exists to solve.
The skill-versioning discipline covered earlier in this blog applies directly to MCP tools, with one added wrinkle — an MCP server's consumers are often external agents you have no visibility into at all.
The skill-discovery-at-scale problem from earlier in this blog gets a new dimension once tools are spread across many independently-run MCP servers instead of one internal registry.
MCP's early examples mostly skip authentication, which is fine for local development and a real gap the moment an MCP server touches production data across multiple users or agents.
Provider-side prompt caching can cut costs substantially, and only for the part of a prompt that's genuinely stable across calls. Structuring prompts to maximize the cacheable portion is a real, underused lever.