Reflection and Memory

Here's one way I think of LLMs: they act like a mirror.

People go into very deep conversations with frontier models. The model can be an incredibly compelling conversation partner, because it'll reflect what you bring, and often shine it up. The vast knowledge-base that's trained into the model lets it go into any number of interesting directions. But watch carefully; often the vibe of the conversation is exactly a mirror of the attitude that you bring.

If you ever got angry with a chatbot because it didn't understand you -- the bot wasn't angry. That was you. The person most affected by your anger? You, too.

I'm cautious in ascribing any measure of "actual intelligence" to LLMs. Instead, I find it useful to approach these interactions as if the model is powerful and intelligent and useful. It will act as a mirror.

Reflection as Practice

Reflection has a second meaning: contemplation, taking stock, reviewing, and considering the events as a whole. This is an essential part of keep: the practice of reflection.

How do we become more skilled? By reflecting on our actions. Repeatedly checking -- before, during and after an action -- to understand whether it was skillful. If we did something unskillful, remember not to do it that way again!

This self-reflection is near the top of several layers; approximately:

1. Write things down. Remember what was said, what was done. (Whether you literally capture everything might vary according to the design of the agentic system). Index it, so that you can search for keywords later.
2. Remember documents and references. A personal agent, or a business agent, will very often encounter documents (results, presentations, reports) and URLs. There might be years of history in emails, scanned records, PDFs. When researching or building, the agent will often discover important resources on the Web. In these cases, you want to remember the "what and where": a summary of the resource, and its location.
3. Analyze for structure and key events. This varies according to the type of interaction. In a conversation session, you will want to remember important requests and commitments. After a meeting, the minutes should include key decisions and stakeholders. When you index a PDF, each chapter might have a different topic. This thematic analysis is different from summarization: it's about finding the meaning rather than the content.
4. Analyze for action. This is true reflection: was the outcome a success? If not, why not? Were there learnings that should be remembered next time? Should the process be changed?

I'll write much more about each of these steps later. They're all important for the function of memory.

Anchorite Lore

Where in this is the agent?

Here's one metaphor that might be useful. There's a medieval English text named the Ancrene Wisse: "Ancrene" means "of the anchoress", and "Wisse" means, approximately, "lore" or "knowledge"; the first version was named Ancrene Riwle, "rule".

An anchorite voluntarily and permanently moved into a tiny cell, walled in. They had a small window into the chapel, and another out into the village. A recluse, but not completely isolated; connected to the world, but separate from it, with a very special role and position.

The Wisse provides guidance to the anchorites. It describes two parts:

Do you now ask what rule you anchoresses should observe? You should by all means, with all your might and all your strength, keep well the inward rule, and for its sake the outward.

The inner rule: how to learn, how to guide one's thought and memory. The outer rule: how to interact with others.

So - a metaphor for an agentic... memory system. But we're building something deeper than recall and relevance. A reflection system. A way to learn.

Try it yourself. Let me know how it goes.