This store is finite. It forgets.
Memories decay as they age, and recalling one reinforces it. Once a memory's value drops below the threshold, the next consolidation cycle deletes it — and the fuller the store gets, the higher that threshold climbs.
The last cycle
Just forgotten
Full log →Content search
Queries the secondary content-search index. Leave the query blank and give just an event id, and you'll get that event's memories instead. Event ids in the results are clickable — they open the whole event.
Results
Create memory
Filter
Memories
Create event
Events whose memories have piled up and gone quiet long enough to be worth condensing, picked out by the most recent sleep cycle. Open one to summarise it.
Events
Forgetting
Every number here comes from the service, computed by the same code that decides what to forget — so what's drawn is what will happen, not an approximation of it. A memory is forgotten once its value falls below the threshold, and the threshold rises with capacity pressure as the store fills.
Decay curve
How a memory of a given effective significance loses value with age. Click the value of any memory in the tables to plot that memory's own curve.
What would be forgotten now
A dry run of the consolidation cycle: it reports what a cycle running now would delete, and deletes nothing. Administrator only.
What was forgotten
The forgotten log: one record per memory a cycle actually deleted,
with the value and threshold that decided it. Optional, and off by
default (consolidation.tombstones.enabled). Records carry no body — this tells you a memory was lost, not
what it said. Emptying the log is administrator only.
What this instance is attached to
An instance knows three things: itself, whatever it dials, and — on a shared database — the peers it finds registered there. Everything else in a deployment (the broker bridges, the ingestor, MCP servers, the CLI) connects to it and is invisible here unless someone says otherwise. So each component below says where it came from: this is what this instance knows, not a survey of what's running.
Components
The configuration of each one, as this instance holds it. Addresses are redacted — no password, key, or connection-string credential reaches this page.
Outbound callbacks
Deliveries waiting to be sent, oldest first. A queue that isn't draining means the receiver is down or refusing: the dispatcher retries with backoff and gives up on a delivery once it's tried too long, and an abandoned delivery is a notification nobody will ever get. Payloads are never shown here — a delivery can carry memory bodies, and this is a view of a backlog, not a second way to read the store.
Storage outside the capacity target
Three tables are deliberately excluded from the store's measured
size: the record of what was deleted must never be the reason live
memories are evicted. So the disk this deployment needs is
consolidation.capacityBytes plus what's here, and
capacity pressure will never reflect it — on the embedded driver
these share the store's own file. Sizes are row counts times a flat
per-row allowance, not measurements.