Firstly five statements:

1. If you start to talk about creating a knowledge sharing culture, or ask for advise on how to get people to share what they know you are missing the point (and you may have missed the boat). In the context of clear need few people will refuse to help someone. However those same helpful people asked by the knowledge management function to share their knowledge/information in anticipation of that need will refuse to do so, or simply fail to cooperate.
2. The issue in lessons learnt processes is not to work out in advance what you might need to know, but rather to ensure that you can connect with people and knowledge objects in the context of need, and that you can make those connections serendipitously. After all, if you already know what you need to know, you will not be looking; if you need to look then you are dealing with uncertainty and do not have a precise idea of what you need, or where you should look.
3. Lessons learnt systems are not about truth, they are about meaning. This contrasts with say process design, where I need truth: I would not want to construct a formal process based on fiction. However in lessons learnt I will not take one solution or one example of past practice. Instead I will naturally gather material from multiple sources, including the might have beens and could have beens.
4. The way people recall the past differs from the actual events. Firstly we only remember what seemed significant enough at the time to imprint on our memory. We may well miss relevant material. We tend to use heuristics in practice, but more structured processes when we reflect on the past. We also tend to combine past experiences, concatenating separate events into single teaching stories. This is valuable and increases both learning and meaning, but it shifts away from accuracy and relevance. The closer to the point of creation the better.
5. We are far more interested in raw material from the coal face or the front line to be less British and more international! The more that material is improved, refined, linked to established practice; the less valuable it is. The process of creating a formal piece of doctrine, or a best practice document involves a process of loss, it also means that the person doing the work overlays their interpretation and the context of the time at which they write. My experience is that raw narrative material survives changes over time, but written documents become dated quickly.

Then four suggested ways forward

1. Capture material at the right level of abstraction, thats not too difficult. If you use an engineering students to capture stories from engineers, then the engineers will naturally tell stories in a form that can be understood by those students. The implication is that you should capture stories a level of abstraction down from the original source. An experienced engineer talking to another experienced engineer talks in code which is highly contextual and may not transfer to others or across time.
2. Get people to record things as they do them, and then index the resulting material either at the time or at the end of the day, so the raw material is interpreted by those who created it. In effect the story teller adds layers of meaning through the indexes which is not present in the original content. This also allows pictures, URLs etc to be used as stories. Cognitive Edge practitioners will know that SenseMaker™ is designed on this basis.
3. Allow people to talk about failure by allowing them to avoid any attribution of blame. There are various ways of doing this. Creating an alternative time line for a project in which you get people to talk about a fictional failure often produces real examples. Telling stories about fictional characters allows people to share actual failures, using archetypes as the characters is even more powerful. This is a bigger subject and I plan a blog on this alone sometime next week when I have reliable internet access.
4. Make capture continuous and a part of the job, not a post job after action review. The more capturing stories is part of the way we do things around here, the better it is. The other advantage is that you can then see trends emerging in the way that people index the story material which allows early intervention. This sort of switch is key to moving from a retrospective and codified set of documents, to a dynamic narrative based learning ecology.