Nick Oak blog
← Back

L-Space Agents: What If Documents Had Souls?

L-Space

It has all started with reading too much Terry Pratchett. This gave birth to an idea of a personal knowledge OS, that I have called pratchett-os. I have built a functional MVP over a weekend. With projects docs, strategy, vision, and many more less prosaic things. And prohibited myself from upgrading this life os for a month. Why? Because I tend to overbuild. That was a proper guardrail. And this writeup of mine is probably an early indicator of another overbuilding spree!

Month has passed and a lot of ideas on how things should rather work have accumulated. MVP has been good - don’t get me wrong. But the usage patterns behind it seems to be “too manual” with a slight micromanagement from my side here and there. And them buzzy fella OpenClaw came in with his interesting concept of memory and bold claim of “bot that remembers everything”.

Proper starvation leads to “building itch” building up. So as soon as month has passed I have started brainstorming. And one of the most crazy ideas looped back to the origins of this OS —> to the L-space itself!

Listen me out

What if every document in my system, every task, every project all of a sudden became the living creature. Each of them with their own goals, ambitions and incentives. Designed in a way that will enable reliable operations of an agentic swarm — becoming 100x enabler for the things I do!

So the project document naturally desires to move forward, To remove all blockers, to become successful business case and thrive.

Experiment document craves for faster resolution of open questions and hypothesis; Energetic Hare per se.

Tech-report from the other side is a cold and rigorous turtle. He wants all claims to be verified against ground truth.He scrutinizes all of the new data before it is added.

Research reports is a spider lair. They weave their nets of connections in order to catch mischievous insects of insights. They weave and connect. They make sure other can use their net in order to catch ‘em all of insights and sparks

This was the initial idea I have been trying to scope. To understand how is that even possible. How to make production ready system from it - not piece of “modern art”. Unfortunately I have been hitting dry claude and the “clicking” wasn’t happening. So I have decided to shelve this idea for some time and let it age for some time.

Eywa

At some point the sudden Avatar rewatch has happened. This has mixed up in my brain with the concept of “document = alive creature” I have been spinning. And gave birth to something beautiful and elegant.

Mycelium network of a whole planet and hyperconnected systems that are caring for each other - Two catalysts of further breakthroughs.

I came out with a concept:

And then:

So the work has divided into two streams

  1. Build mycelium network of previous Claude Code sessions in order to allow for the seamless context transfers for new Claude Code sessions and some subagents
  2. Normalize and cleanup documents - proper spring cleaning; based on the cleanup meta observations - build guardian agents for each domain

And them two merging into

  The Vision (from experiment doc)

  Sessions that know what came before (context continuity)
  Memory that flows from sessions without manual discipline
  Ghost agents that sip context from the river of episodes

So far

The work on mycelium converged to a mcp that does:

  1. Raw Claude Code session json —> deterministic markdown with full user messages, claude messages, files read and written, subagents briefs, tools used (succinct); So basically much more stable for further LLM indexing: session.md
  2. Session.md —> digest.md & digest.json—> LLM generated “executive summary card of the session”; markdown file to be fed into context; json file for search and routing. Highlight of decisions taken, assumptions made, etc
  3. MCP server that serves these cards on request during session start and for mid-session grounding of “what have we been doing after all?”

Guardian is still WIP, because imagine explaining to dry claude what it is that you exactly trying to build. Though my bet that dry claude thingy- is due to the quantization that they have on their servers. higher precision = wet claude. lower = dry

P.S

there is another realization of mine that it is okay to post things that are work in progress. So let’s treat this post as a beautiful draft of several ideas, that I hope will converge to a reasonable tools, that will do the expected 10x for my local knowledge management system

As per now this post is as explicitly open loop. It is waiting when I will finish River and Guardian agents and return here, or to the next post with a proper feedback. And break down of things that have worked

P.P.S

this is the mental task I am giving claude several times per session. Worth trying to give myself the same prompt and just dump here what are my assumptions towards things that I am building

Unstructured, as they should be:

  1. Concept of a “document = living creature with its own incentives” combined with “with a proper incentives design such system will work for me” - pure speculation. It’s rather belief, not yet tested
  2. Guardian lobbying “documents - stakeholders”, as an additional complexity layer to tricky enough already layer (1) — will make this somehow working
  3. River protocol will work better than claude code session leaving a handoff file at a random location.
  4. River protocol will NOT lead to slop slopping around

P.P.P.S

So this has expectedly exploded complexity wise. So river became just the “automatic handoff protocol with simplest possible routing” in order to become useful. And concept with guardians rather became get-knowledge and put-knowledge. But without initial “wide open mind plan” this wouldn’t be possible.