WEBVTT

1
00:00:08.570 --> 00:00:13.828
When it comes to AI, memory is one of
the most essential, and least understood,

2
00:00:13.828 --> 00:00:16.677
ingredients of intelligent applications.

3
00:00:17.513 --> 00:00:23.560
Most people imagine memory as a static
database: information goes in, and later,

4
00:00:23.560 --> 00:00:26.096
if you're lucky, it comes back out.

5
00:00:26.453 --> 00:00:30.524
But inside ServiceNow,
memory isn't a bolt-on.

6
00:00:30.881 --> 00:00:36.030
It's a dynamic system embedded into the AI
framework, so agents can learn,

7
00:00:36.030 --> 00:00:39.504
recall and personalise
interactions across time.

8
00:00:40.340 --> 00:00:44.520
Memory is what transforms
a chat into a relationship.

9
00:00:45.740 --> 00:00:49.082
At the heart of ServiceNow's
memory system are three layers

10
00:00:49.082 --> 00:00:51.795
that mirror how humans think and recall.

11
00:00:52.632 --> 00:00:56.730
Short-term memory is the context
within a conversation.

12
00:00:57.087 --> 00:01:01.811
It helps the agent keep track of the
current flow and respond coherently.

13
00:01:02.168 --> 00:01:06.651
Episodic memory is the historical
record of interactions and events:

14
00:01:06.651 --> 00:01:09.377
chats, incidents and actions.

15
00:01:09.734 --> 00:01:12.357
And long-term memory
holds the enduring facts

16
00:01:12.357 --> 00:01:17.284
and preferences that persist across
sessions, linked to individual users.

17
00:01:18.121 --> 00:01:22.686
Long-term memory lets an agent say,
“Welcome back, I remember your last

18
00:01:22.686 --> 00:01:26.703
laptop issue,” without ever needing
a case history in front of it.

19
00:01:27.923 --> 00:01:31.831
Behind the scenes,
this runs on a clean schema.

20
00:01:32.188 --> 00:01:36.830
Memory categories define what types
of information can be remembered.

21
00:01:37.187 --> 00:01:42.372
The agent-category mapping controls
which agents can access which memories.

22
00:01:42.729 --> 00:01:48.390
And the memory table stores the individual
facts, tied to users and categories.

23
00:01:49.227 --> 00:01:54.766
Each memory has a relevance score, which
increases when it's recalled successfully.

24
00:01:55.123 --> 00:01:59.167
That teaches the system what's
most valuable to remember.

25
00:02:00.387 --> 00:02:04.911
When a user mentions something memorable,
like “I prefer Teams calls”,

26
00:02:04.911 --> 00:02:08.208
the agent checks whether
it's allowed to store that fact.

27
00:02:08.565 --> 00:02:12.283
If it is, it saves it
under the right category.

28
00:02:13.120 --> 00:02:14.520
In a future session,

29
00:02:14.520 --> 00:02:18.256
when the context calls for it,
the agent retrieves the fact,

30
00:02:18.256 --> 00:02:21.811
injects it into its reasoning,
and responds accordingly.

31
00:02:22.648 --> 00:02:24.455
The more a memory is used,

32
00:02:24.455 --> 00:02:29.749
the stronger its relevance becomes, which
keeps the system personal and efficient.

33
00:02:30.969 --> 00:02:35.339
Once you start customising
memory, it gets interesting.

34
00:02:35.696 --> 00:02:39.754
Each layer offers a place
for intelligence and control.

35
00:02:40.591 --> 00:02:45.124
The agent layer defines
when an agent should remember or forget.

36
00:02:45.481 --> 00:02:49.000
A conversational agent
might store preferences,

37
00:02:49.000 --> 00:02:52.650
and a troubleshooting agent
might remember configurations.

38
00:02:53.007 --> 00:02:56.838
The retrieval layer decides
how memories are fetched and ranked,

39
00:02:56.838 --> 00:03:00.026
using relevance scores or recency filters.

40
00:03:00.383 --> 00:03:04.106
The governance layer adds
privacy rules, retention limits

41
00:03:04.106 --> 00:03:08.368
and user-controlled deletion,
to balance usefulness with trust.

42
00:03:08.724 --> 00:03:13.181
And the prompt layer decides how and
when memories surface in the conversation,

43
00:03:13.181 --> 00:03:16.301
from “Last time you
said” to silent context.

44
00:03:17.521 --> 00:03:20.071
Memory isn't just a feature.

45
00:03:20.427 --> 00:03:25.817
It's the connective tissue between
conversations, data and experience.

46
00:03:26.174 --> 00:03:30.385
In ServiceNow, it's what lets
AI agents evolve from reactive

47
00:03:30.385 --> 00:03:33.180
assistants to proactive collaborators.

48
00:03:33.536 --> 00:03:35.710
Combine memory with orchestration,

49
00:03:35.710 --> 00:03:40.398
retrieval and reasoning, and you get
systems that know what's happening now,

50
00:03:40.398 --> 00:03:42.499
and remember what happened before.

51
00:03:42.856 --> 00:03:46.615
That continuity is the essence
of intelligence.

52
00:03:47.452 --> 00:03:50.545
Memory in ServiceNow isn't magic.

53
00:03:50.901 --> 00:03:55.204
It's built from tables, like any
other feature in the platform.

54
00:03:55.561 --> 00:03:57.911
As AI agents mature,

55
00:03:57.911 --> 00:04:02.268
memory will define their personalities
as much as their capabilities.

56
00:04:02.625 --> 00:04:06.927
The so-called smartest systems
won't just generate answers.

57
00:04:07.284 --> 00:04:10.282
They'll remember why those answers matter.
