How is this different/better than SOUL.md?

[TomLucidor]

Check the following, there might be ways to make the two work together well?

[ohad6k]

good question, they're close cousins with different raw material.

soul.md builds a persona from stuff you published: tweets, essays, posts, notion notes. that's you performing for an audience, and the output is pointed at making the agent write and sound like you (identity, opinions, style).

ditto only reads your raw coding session logs, the .jsonl sitting in ~/.claude and ~/.codex. nobody performs in there, you're just working. so what comes out is a working profile instead of a voice: what code you reject, what 'done' means to you, when you ask for proof, where you get stuck. the agent doesn't learn to sound like you, it learns to work with you.

and yeah they'd stack well. run ditto on your logs, then feed the you.md in as one more source for your soul.md if you want the voice side too. an adapter that exports ditto output in soul.md's expected shape would be a small lift, leaving this open to track it.

[TomLucidor]

@ohad6k thanks for the response. Here is another tricky question worth documenting: how can this work with llm-wiki or obsidian or whatever graph memory there are? Or maybe skill subfolders? What about borg or other skill-merging tools? https://github.com/aplaceforallmystuff/claude-agent-borg

[ohad6k]

intersting questions, worth documenting, adding a compatibility note to the readme.

different layers. graph memory (obsidian, llm-wiki) stores what you KNOW. ditto outputs how you WORK. One lean profile loaded on every task. they stack: the you.md is plain markdown, drop it in your vault, link it in your graph, wherever your stuff lives.

one deliberate limit: ditto won't mine your notes as source. notes are curated you, sessions are raw you. it only mines the raw logs.

skill folders + borg: you.md installs as a completely standard skill, so any loader or merger handles it. fun caveat for Borg, it adapts skills to your conventions, and you.md IS your conventions. it's the one skill borg should absorb verbatim.

[TomLucidor]

Could you doc something like a unified system as an example for documentation? Feels like this can go somewhere more clean. Notes could be less like a curated self but also a transcript-of-thought or designer self rather than a project-manager self from Ditto.

[ohad6k]

did the thing, you named a real gap so i built it out.

you were right that you.md is the manager self, laws and rules, because that's what the reducer asks for. so i added lenses: same logs, different reducer, different self. there's a thinking self (how i reason and debug) and a designer self (how i make taste calls) now, not just the working one.

all three are in the repo mined from my own 8 months, so you can see one real person cut three ways:

• working: examples/lenses/you-working.md
• thinking: examples/lenses/you-thinking.md
• designer: examples/lenses/you-designer.md

and the unified system you asked for, one picture: examples/unified-system.md. graph memory = what you know, ditto = how you work, soul.md = how you sound, and how they stack.

thanks for pushing on this, it made the tool better.

[TomLucidor]

I almost forget synthesis and versioning, seems like those are important points. Also skills/hooks/workflows (not gonna ask for weekly update loops, most likely structured skills that solicits from users to "spill the beans" more easily) @aplaceforallmystuff is somewhat inspiring and I wonder why Ponytail, Caveman, Anti-slop, dialectic, "mental frameworks" etc cannot be their own sub-project or subjects of study.

[ohad6k]

your ideas are strong, this is a real direction i want to work on, not just notes.

versioning is already the plan, the --diff on the roadmap, re-mine and show what got stronger and what you dropped. the spill the beans pass is the sharp one, it asks you the few things the logs can't show and folds it in, adding it now.

and ponytail/caveman/anti-slop as their own studies, that's the fun part, named ways of working while ditto mines the unnamed one that's already yours. composing them is where it gets real. you keep naming the actual gaps, let's build toward it.

[aplaceforallmystuff]

Hi folks, got tagged so popping my head up. 👋🏻

A minor quibble on soul.md, from my perspective and how I've implemented it into cerebro: it's actually about defining an AI's identity — its values, boundaries, and relationship with humans. It's not about mining your public content; it's about capturing the emergent identity of the AI itself through your working relationship with it.

I like this ditto concept because it appears to be a complementary system to defining the operator's identity. I have a similar 'log-to-daily' skill that covers this for me, and writes my outputs to my Obsidian vault. So I'm building up a memory of how I work with Claude Code with the vault, qmd, claude-memetc.

So how I see it:

• graph memory = what you know
• ditto = how you work
• voice/persona tools = how you sound
• soul.md = who the agent is

which makes ditto and soul.md near mirror-images - one profiles the human from the logs, the other lets the agent articulate itself.

Because Cerebro is more of a 'second brain' with agentic capabilities, the agent-borg mentioned above helps to integrate other Skills, Agents, Gems etc that I think might be useful - but by absorbing them into my own tooling, instead of copying blindly, so that it's adaptive and not just whacked on top.

Might just borg ditto ;-)

[ohad6k]

appreciate the correction, you built it so you'd know. i had soul.md wrong and just fixed it in the doc.

your layering is cleaner than mine:

• graph memory = what you know
• ditto = how you work
• voice/persona = how you sound
• soul.md = who the agent is

the mirror-image line is what clicked for me, ditto profiles the human from the logs, soul.md lets the agent articulate itself. updated the unified-system doc to say exactly that and credited you.

please do borg it, that's the whole point. it installs as a plain skill so it should absorb clean. would love to see how it sits inside cerebro.

[TomLucidor]

Another set of ideas I have in my head would be the whole Graph-of-Thought or Atomic thoughts idea, since some people have split identities when problem-solving, maybe semantic search skills/abilities to handle that as well. "Memory palace" feels like an abused concept but maybe it can be rectified?

Bonus: relationship tracker https://github.com/SivaRamSV/paaw

[ohad6k]

The atomic-thoughts angle is the sharp one. Right now ditto flattens you into one profile, and the split-identity thing you're describing is exactly what gets lost, the mode you drop into debugging vs the one you make taste calls in. I already keep contradictions as evidence instead of resolving them into one voice, this just pushes it further.

How I'd do it without blowing up the tool: the reducer emits linkable atoms. Each trait gets a stable id plus links to its receipts and to related traits, and you.md becomes a flattened view of that graph. Graph-ready for free, and semantic-search-per-task later is a loader swap, not a rebuild. No graph db needed yet.

Full Graph-of-Thought + semantic retrieval is a v2 direction, not the next release. That one stays bounded and local on purpose, a retrieval layer now breaks that.

paaw is a good pull, but note it's Postgres + Apache AGE, a whole backend. "Model the people around you" is a different product and a different privacy surface than "model you", so I'd keep the relationship tracker as its own thing, not folded into ditto.

Memory palace I'm parking. You already called it an abused concept and I don't see the real version yet either.

Banking atoms-with-links now, the rest goes on the list. You keep finding the actual gaps.

[ohad6k]

Tom, if you got any ideas or just want to discuss more about ditto you always invited to the ditto discord : https://discord.gg/VNnMq2U5r

[TomLucidor]

Other than how "paaw" can be turned into agentic tulpamancy (less privacy issue and more "studio simulator" with public blogger dialogue), the graph thing could ideally be either some kind of SQLite based graph system OR just rely on fzf or some other terminal tools for traversal. Semantics will need embedding (from the same inference engine so MLX have rapid-mlx and GPUs have vllm usually + vector DB (again ideally SQLite based).
Will probably back-and-forth between Discord and GitHub discussions for more ideas. PKM, ZK, BASB, Digital Gardens, LLM-wiki, Toulmin links/edges, ... there are a lot more well-organized way to represent mental associations.

[ohad6k]

SQLite-based graph is the right call. It keeps the whole thing local and backend-free, one file next to your profile, which is the line I don't want to cross (paaw is great but it's Postgres + Apache AGE, a real server). Atoms as nodes, receipts as edges, stored in one .db, is basically the atoms-with-links idea with a real query layer on top. fzf for traversal is a fun cheap first version before any of that.

Embeddings I'd keep optional and local, exactly like you said, same inference engine so nothing leaves the machine. That's the v2 semantic-retrieval loader, not v1.

The Toulmin framing is the one that clicked for me. Ditto already separates the claim (the installed rule) from the grounds (the verbatim receipts), so the edges are half-built already. Warrant and backing just name the link type.

Tulpamancy / public studio-simulator is a different animal. That's personas performing outward, and ditto points inward at how you actually work. Interesting on its own, but not the same product.

v1 stays lean and flat. The graph is where this goes next. Keep the ideas coming, this is the good part.