I was talking recently to a colleague, John Hutchinson, who is thinking about ways to augment our memories and assist us in linking this to external knowledge. There have already been a number of research projects in this sort of area (e.g. automatic diaries based in active badges at Xerox in Cambridge, MIT cyborgs) and there are potential practical benefits.

This led me too wonder just how much memory it would take if we really did try to remember everything in our lives. Now I recall, from articles many years ago, that human long term memory is only laid down at about half a bit per second, but we are very good (and sometime bad) at selectively remembering things that are significant. A computer assistant would be able to remember more than this, but would be far less selective, let's say it remembers everything we see or hear.

With good compression it is possible to get pretty high quality video and audio within ISDN capacity, so let's say 100 kbits/second will be needed. At a biblical life span of three score and ten, our lives are a mere two billion seconds. So our complete life experiences amount to a 200 terabits, or about 400 hard disk drives.

after T years: storage size, S = aT 2-T/b where a = data/year = 300 billion bytes b = halving time = 1.5 years   differentiate to find maximum: dS/dT = a (1 - T loge2/b) 2-T/b   So worst size at T = b/loge2 that is about 2 years

In fact, we have seen a constant reduction in cost and size of memory over past years. I'm not sure of the exact figures, but let's take a Moore's law figure of doubling every 18 months. At this rate, the real crunch time, in terms of worst memory requirements are about 2 years in (see side bar). After this point the increases in storage capacity would mean that the accumulated memories would take less and less memory each successive year.

After two years we would have accumulated around 60 million seconds of data, that is around 6 terabits. If we started such a system now (and assuming the increase in storage capacity) this would require about half a dozen hard disks over the coming 2 years, but we could discard these for smaller, cheaper, larger, faster disks over coming years. At first we would need to use wireless technology and remote storage, but as storage devices shrink, we would be able to carry our lifetime memory with us. Each year they would shrink.

If trends continued, in 70 years storage capacity will have doubled around 47 times - that is capacities will increase 100 trillion fold. At this rate 200 terabits will only take the space of 2 bits today! There are quantum limits to information density, so let's look at this from this perspective. Let's assume we need 1000 atoms of silicon to store a single bit without quantum problems (pretty conservative). 200 terabits would require a device containing 2x10¹⁷ atoms of silicon. This is about 10 micrograms or a sphere less than 200 microns in diameter.

So, if we did this for a baby born tomorrow when the baby dies in 2072, the storage device needed for a whole life's experience would, be the size of a small grain of dust.

Related links

If you know of further work in this area please let me know.

John Hutchinson, will be continuing to work on Digital-Me at Lancaster.

Memories for Life is one of the UK Grand Challenges for Computing Science. Ehud Reiter has a Memories for Life resources page with lots of links to the challenge and related material. There is also a corresponding track of the UK Foresight Commission called Mnemo Net.

Bradley Rhodes worked at MIT on Remembrance Agent, which watches everything the user types (in emacs!) and prompts with related things in the past. He is now in California at Ricoh Innovations, where he's continuing to work in related areas.

There have been a number of projects at XRCE Cambridge on augmented memory. These include Forget-me-not a PDA based ubiquitous memory aide and Pepys, which used active badges to create automatic 'diaries' and was later augmented by a Video Diary.

The Factoid project at Compaq Western research Lab (see Technical Note TN-60) imagined small ubiqutous devices that would capture many small 'facts' per day, each only a few hundred bytes in size (business card, temperature etc.). Based on 1000 facts a day they caclulated 73Mb/year and so "can be saved forever and constitute a sort of history of the user's life".

The Sentient Computing Project at AT&T Laboratories Cambridge uses a range of sensors to give the computer a model of people and what they are doing. The primary aim of this is to make interaction more fluid becasue users and computer syustems would share a common model of the world. however a side effects is to allow users to be able to view this recorded history. Each user has an `information hopper', which can be displayed as a timeline (see example)

Kasim Rehman at Cambridge Computer Laboratory made a survey of '101 ubiquitous applications' including a number of personal information management applications, all with web links.

Gordan Bell at Microsoft has been recording all the paper, emails, faxes he has seen or produced as well as videos of presentations etc. He's reported issues surrounding this in various of his publications, for example Storage and Media in the Future When you Store Everything and makes estimates of the storage capacity required - he estimates a gigabyte a year (note this is only formal materials hence a lot less than the estimates in this article). This project used to be called Cyber All, but has been renamed MyMainBrain .. hmm.

In a project called the I Sensed Series, Brian Clarkson at MIT has built a wearable computer that records video (including looking backwards), audio and orienetation data for up to 10 hours a day. He uses a 10 gigabyte disk to record up to 2 days data. This is a larger amount of storage per day than projected above, I;d guess becasue it is largely uncompressed.His research also includes other projects the 'familiar' and the 'auto-diary' in the same area.

Indexing is a critical issue for lifetime information. The MEMOIRS project in Loughborough in the early 90s involved recording all documents, faxes etc. and annotating them with users' own definable tags, colour coding etc. in order to index them. More recently David Gelernter's group at Yale worked on another timeline based storage system called Lifestreams (Wired article). This has now been incorporated into a commercial product Scopeware. At Toronto University a PhD student Byron Long is working on another time-based indexing system called TimeStore. Also at Xerox PARC,the Placeless Documents Project looked at attribute-based retrieval of docunments.

... and a vaugely related issue ... WHY don't file open dialogue boxes on PC and MAC show the currently open windows as a special category in "save file" dialogue boxes? Surely this is often the most likely place to store something!!!

To put this kind of technology in context look at Bonnie Nardi's position paper on nanotechnology.