Every decision, lesson, and trace is recorded as it happens, not reconstructed after the fact. The record is a byproduct of the work, not a chore on top of it.
remember · living memory
remember is a durable knowledge layer for AI agents. It holds what they learn, retrieves what is relevant, and curates the corpus over time, so the next session starts from everything the last one knew.
What it is
Agents are brilliant and amnesiac. Every session starts from zero, re-learning what the last one already knew. remember is the layer that holds on: three primitives that turn ephemeral context into durable institutional memory.
Every decision, lesson, and trace is recorded as it happens, not reconstructed after the fact. The record is a byproduct of the work, not a chore on top of it.
Vector and semantic search, scoped to the agent and the task at hand. The right context arrives without flooding the window with everything ever learned.
Dedupe, supersede, and retire. Knowledge that is wrong or stale does not just accumulate. The corpus stays small enough to trust and current enough to use.
SLF is the interface every read and write goes through. Substrate, Lens, Frame: it renders the slice of memory an agent needs, scoped and accountable, instead of handing back the whole store.
render(substrate, lens, frame) → slice The SLF design →The compounding loop
Memory is only worth keeping if it comes back at the right moment. remember closes the loop: what one agent learns, the next one inherits.
Agents emit decisions, lessons, and traces as they work. Nothing has to be remembered by hand.
Raw events become structured, searchable knowledge: who decided what, and why.
The next session pulls what is relevant, scoped by lens, before it starts cold.
Conflicts resolve, duplicates merge, and outdated knowledge cools rather than vanishes. The record stays honest.
The loop runs every session. Knowledge does not just pile up, it compounds.
Kept, not lost
The commitments
Memory that you cannot trust is worse than none. remember is built on four commitments that keep the record worth reading.
Commitment 01
Durable by default
Knowledge persists across context limits, sessions, and handoffs. Agents do not start cold, and nothing learned is lost when a window closes.
Commitment 02
Scoped, not dumped
Retrieval is filtered by lens and role, so an agent gets the slice it needs. Memory that floods the context is as useless as no memory at all.
Commitment 03
Curated, not hoarded
The hard part of memory is deciding what to keep, not just what to add. remember dedupes, supersedes outdated facts with corrections, and reinforces what's proven valuable over time. Aged knowledge cools rather than vanishes — deletion is reserved for what's actually wrong, not what's gone quiet.
Commitment 04
Provenance, always
Every piece of knowledge carries where it came from and when. You can trace a belief back to the decision that set it, and revise it when the world changes.
The pair
remember stands on its own, behind any agent system. It is also one half of a pair: the engine that makes the work, and the memory that keeps what the work taught.
Plans, builds, tests, and ships through governed gates. Produces the work, and a record of exactly how it was made.
Holds what every run learns, and serves it back to the next one. Turns a thousand sessions into one accumulating body of knowledge.
create makes the work. remember keeps what it taught.
Get in touch
Reach out about putting remember behind your own agents, on its own or paired with create.