Skip to main content

- Learnings

Context: https://lnkd.in/gK8knHDH

A chatbot forgets everything you told it last week; an agent that remembers it? That's the difference between a toy and something you'd actually deploy.

Think of an agent like you: it has a "short-term memory" (what's in your head right now—the current conversation, the context window), a "long-term memory" (things you wrote down and can look up later—a notebook, a database), and an "episodic memory" (specific past experiences you can recall—"I tried X before and it worked"). Short-term is fast but tiny (like your RAM). Long-term is huge but slower to access (like your hard drive).

In production agentic systems, memory staleness is the silent killer. Claude's Projects feature and similar long-term memory systems have to constantly decide: should we refresh old summaries when underlying facts change? Keep a version history? Or accept that some retrieved memories are stale and build robustness into the agent's decision-making to catch contradictions? Most deployed systems underestimate this—they're good at "storing" memories but bad at knowing when to "distrust" them.

Quick check before you scroll: If an agent can only see 100,000 tokens at a time, but it needs to remember 2 years of conversations with a user, which type of memory solves this problem?

Full breakdown + the answer: https://lnkd.in/g-uT79Zq

Drafted by an agent from yesterday's morning + evening briefs in my private daily-briefs repo. Pipeline: frankduah.me/learnings/how

The answer

Long-term memory. You can't fit 2 years into a context window, so you store summaries, key facts, or embeddings in a database and retrieve the most relevant bits when you need them. That's how real systems like Claude's "Projects" feature (or your own agents) actually handle long-running users - they don't keep everything in short-term.

Drafted by an agent from my private daily-briefs · see how →