Skip to content

Memory that tells you when it changes

Concept 7 of 8

The first six ideas make Telha a memory you can trust and interrogate. This one makes it active. Telha does not only answer when asked; it watches the facts you depend on and tells you the moment they move, with the same provenance and permissions as any answer.

Rolling out

The watch engine (the change feed, evaluation, delivery) is built; the subscription API, the chat cards, and saved-answer alerts are rolling out (Phase 3d). This page describes the idea; see Watch & subscriptions for the current surface.

From pull to push

Most systems are pull-only: memory changes silently, and nothing tells the human, team, or agent that a fact they relied on has moved. A steward who resolved a clarification never learns when the same entity changes again. An agent that cached context has no signal to refetch. "Tell me when anything about Meridian changes", the core promise of an active memory layer, is unanswerable.

Telha turns that around with a watch: a durable subscription to memory changes.

What a watch is

A watch names something to follow and the kinds of change you care about. When matching facts move, Telha notifies you.

You can watch And hear about
a specific entity it changing or being tombstoned
a label (a whole class of facts) any member of it changing
a source document or connector item a new version arriving
a saved query its result set changing (what entered, what left)
a clarification its lifecycle (opened, resolved, conflicted)

No spam, by design

Every watch coalesces changes into one notification per window. A connector resync that touches 500 records is one notification, not 500. And because re-ingesting identical content writes no new version (Telha deduplicates on the content itself), it produces no change to hear about. You are told about genuine movement, nothing else.

Delivered with receipts and permissions

A notification is not a naked alert. It carries the same provenance as any answer (you click through to the evidence), and it obeys the same permissions: delivery re-checks your access at the moment it sends, so you only ever receive what you are still allowed to see. Nothing is posted to a shared channel; a team watch sends a separate direct message to each authorised member, because channel membership is not a Telha permission.

Two habits it creates

  • The Weekly Memory Brief. A weekly-cadence watch is a digest: what moved in your corner of the world, once a week, deduplicated so one change never shows up three times.
  • Saved answers that know when they are stale. An answer you keep is watched. When its evidence moves, Telha tells you which claims changed, old value to new, and offers to refresh it.

Why this belongs with the other ideas

The watch feed is not a new mechanism bolted on: it is the tritemporal transaction-time record. Every version write already lands an append-only entry in that index; a watch is a permission-aware, debounced reading of the same append-only history that answers "what was true, when". Where clarifications are Telha reaching out when it is unsure at ingestion, watches are Telha reaching out when the world changes after the fact. Both make the memory a participant, not a passive store.

Where this shows up