LECE: How Founder OS Learns From Your Life

Every serious agent logs everything. Tool calls, timings, failures, final answers - JSONL traces piling up in a folder like security footage nobody watches.

Founder OS was the same for months. Rich flight recorder. Zero learning loop. The system could act but not grow. It was a capable body with no memory that compounded.

I wrote about why I built the body in the first place - five thousand outreach threads, a lying spreadsheet, Telegram as command center: I Built Founder OS Because I Was Drowning in Five Thousand Conversations. This post is what I bolted on top when RAG and CRM were not enough.

It is called LECE - the Lived-Experience Cognitive Engine. The full technical whitepaper is embedded at the bottom of this post (PDF). Source and implementation live in the Founder OS repo. I will tell you what it does, why I built it this way, and what I think is actually new.

The Thesis in One Sentence

The most valuable training data for your assistant is not the internet. It is the private, day-by-day record of your company - and almost nobody uses it because nobody closed the loop on-device without shipping your life to a cloud trainer.

LECE closes the loop.

Three Pillars (Not Three Features)

LECE is not a module you toggle for “smarter replies.” It is three reinforcing systems:

flowchart TB
    subgraph live [Every day you work]
        act[You + Founder OS act]
        trace[(Flight recorder traces)]
    end

    subgraph p1 [Pillar 1: Distillation]
        episodes[Score episodes]
        principles[(Principles + manual)]
    end

    subgraph p2 [Pillar 2: Preplay]
        twin[Digital twin of your business]
        sandbox[Sandbox - no real sends]
        preplay[Rehearse before commit]
    end

    subgraph p3 [Pillar 3: Workspace + Trust]
        ws[Global Workspace attention]
        trust[Earned autonomy per action]
    end

    act --> trace
    trace --> episodes --> principles
    principles --> act
    act --> preplay --> sandbox
    preplay --> twin
    twin --> preplay
    ws --> act
    trust --> act
    act --> twin

Each pillar echoes published research (EvolveR, continual experience internalization, ProPlay-style world models, Global Workspace architectures). The contribution is not inventing those ideas in isolation. It is wiring them into one circuit on one founder’s private data.

Pillar 1: Self-Distillation - The Diary Someone Finally Reads

Every turn Founder OS lives gets appended to data/traces/*.jsonl. In most agent projects that file is for debugging. Post-mortems. Git blame for the model.

LECE reads it on a schedule - nightly, after memory consolidation - and runs a pipeline I care about deeply:

1. Segment and score episodes (0–1) from signals already in the system: did verification pass? did I approve the action? did outreach succeed? did self-healing kick in mid-task?
2. Off-policy filter - only keep genuinely good episodes as teaching material. Learning from your current flailing reinforces mistakes. The 2026 literature on continual internalization is brutal about this. I listened.
3. Distill principles, not anecdotes - not “emailed Sarah Tuesday” but “for technical-founder outreach, lead with a specific product observation before generic praise.”
4. Deduplicate and decay - near-duplicate principles reinforce; stale advice quarantines. A mind that only accumulates gets cluttered. LECE forgets on purpose.
5. Inject - principles land in a living Personalized Operating Manual, surfaced turn-by-turn and step-by-step during reasoning, not dumped once into a bloated system prompt.

After a few weeks the agent reasons differently than day one - not because weights changed (there is an optional local LoRA path for that later), but because it carries compressed wisdom from our lived experience.

And it never left my laptop.

That matters when your CRM contains investors who ghosted, enterprise pilots that stalled, and the exact wording you used when you were wrong about pricing.

Pillar 2: Digital Twin + Preplay - Rehearse Before You Send

Distillation makes the system wiser about the past. The dangerous moments are irreversible: the email to your best lead, the public post, the calendar commit you cannot unsend.

Humans rehearse. “If I open aggressive, he balks. If I anchor softer and follow up Thursday, he engages.” LECE gives Founder OS that habit.

Digital twin - structured state of my business: CRM pipeline by stage, tasks, goals, runway, recent outcomes. Layered on top: a procedure graph learned from the action log - when the business looked like this and I took that action, this tended to happen, with this reliability.

Sandbox - when simulation mode is on, a flag propagates through execution context and the tool layer intercepts every real-world side effect. send_email in the sandbox does not send. It returns a prediction. Enforced at the registry level so rehearsal cannot leak into production. I can let the agent imagine freely; the world does not feel it.

Preplay - before high-stakes tools fire, Founder OS generates candidate variants, estimates outcomes (procedure graph, swarm role-play, or hybrid), picks a branch, and only then routes the real action through the normal approval gate - with reasoning attached so I see why it chose what it chose.

After the real action lands, expectation feedback compares prediction to reality and updates transition reliability. Imagination gets calibrated. Wrong rehearsals hurt the model’s self-trust the same way wrong predictions hurt yours.

When I was juggling Stamped Energy conversations with plant managers who have seen a hundred vendors lie about OPC-UA integration, I did not want an agent that drafts fast. I wanted one that thinks before it spends my credibility.

Pillar 3: Global Workspace + Trust You Earn

The third pillar is what makes it feel less like a tool and more like something awake - without handing it keys to your kingdom on day one.

Global Workspace (in the cognitive-architecture sense): specialized processes bid for attention on a blackboard - pending approvals, overdue follow-ups, goal pressure, recent failures, twin risk when runway or backlog goes red. Highest salience wins and gets broadcast into deliberation. Optional continuous mode ticks this every couple of minutes so the system can tap your shoulder before you remember to ask.

But proactive agents are terrifying unless autonomy is earned.

Most products give you a global dial: cautious, balanced, yolo. That is nonsense. You trust a cofounder to book travel before you trust them to email your lead investor.

LECE tracks per-action-type competence from outcomes and the failure ledger. Autonomy scales with demonstrated success, modeled risk tolerance, and preplay danger signals. It can grant more caution than static settings. It can never grant less than the hard floor - irreversible actions still hit the approval gate. Sacred list. Non-negotiable.

There is also a lightweight theory of mind of me - communication style, decision speed, risk appetite - refined from how I actually behave, not a onboarding form I lied on.

A Composite Day (All Three Pillars)

3 a.m. Distillation runs. Yesterday’s successful outreach episodes become a reinforced principle about technical-founder messaging. A stale principle about a channel I abandoned decays into quarantine.

8 a.m. Briefing surfaces three follow-ups I would have missed.

10 a.m. I say: follow up with the lead from the energy conference. The agent recalls my terse style, preplays three draft variants against the twin of that relationship, picks the observation-led one, routes it to my phone for one-tap approve. I tap yes.

Four days later the lead replies. CRM updates. Tonight that episode becomes tomorrow’s teaching material. The procedure graph ticks reliability upward for that kind of message in that kind of state.

That is the flywheel: act → record → distill → imagine → act better → measure → distill again.

Not a demo. A loop.

What Is Actually New (Intellectual Honesty)

I did not invent self-distillation, world models, or Global Workspace theory. I stand on EvolveR, EvoSC, continual experience internalization, ProPlay, and the 2026 GWT agent papers - cited in the repo.

What I claim:

1. The problem formulation - personal, private, longitudinal, on-device improvement from one human’s operational life. Not anonymous benchmarks.
2. Closed loop on private life - traces → principles → behavior; twin → preplay → safer actions → measured outcomes → sharper twin. Integrated in shipping code.
3. Sandbox at the tool registry - structural guarantee that rehearsal cannot touch production.
4. Trust as a learned, per-capability quantity with a hard safety floor.

Synthesis counts. The transistor was known physics. The arrangement was not.

Config and Limits (Because I Am Not Selling Magic)

LECE is mostly default-on for safe learning (Tier A distillation). Heavy stuff is behind flags: continuous workspace loop, Tier B LoRA training on local GPU, swarm preplay rollouts that are LLM-estimated not full simulation.

Read the config table in the whitepaper below or docs/LECE.md in the repo if you are running it yourself. Start with distillation and preplay. Turn on continuous cognition when you accept the API bill.

Swarm preplay is not a physics engine. Founder theory-of-mind is heuristic until you give it months of data. Tier B weight training needs a GPU and discipline about eval gates.

I would rather ship honest limits than a render.

How This Connects Back

Founder OS is the body - tools, CRM, Telegram, swarm, immune system. That story is here.

LECE is the part that remembers - the thing that turns a sequence of brilliant disconnected moments into something that compounds.

If you are building agents for founders (or for anyone with a longitudinal, high-stakes life), stop optimizing only for benchmark cleverness. Optimize for continuity: what happened, what worked, what to rehearse, what to trust.

The model is the engine. LECE is what keeps it from resetting every morning.

Code and docs: github.com/officiallyutso/Founder-OS. If you build on it or think I am wrong about any of this, tell me. The machine is listening. For once, it will remember what you said.