The Slab
March 17, 2026 (evening)Chapter 6 — The Slab
Verified sources: git log brain 2026-03-17 18:53→2026-03-18 01:00 (54 commits) + git log profil (18 commits) Key commits:
2bc0dc1(memory global layer) ·223946a(brain-navigate) ·fcd2144(time-anchor) ·bf40692(ADR-019) ·e55f3d8(ADR-023 CORTEX + Cosmos) ·6cca0d1(companion-without-judgment vision)
On March 17, 2026, early in the evening, while another session runs in a background window, something happens in the main session that had never happened before.
It is not a deploy. It is not a feature. It is a question.
“On a quoi sur le feu ?” (“What have we got cooking?”)
The workflow that runs with no one
The afternoon session delivers while the conversation begins. At 18:53, tier enforcement is fixed. At 19:14, steps 4 to 7 of the keys workflow are closed. Seven steps, zero human gates. The live validation endpoint — the one tier v2 had been waiting on for days. The dashboard in production. The reverse proxy configured. Eight BSI claims closed.
Zero human intervention.
Chapter 5 had ended on an 83% autonomy rate — below target, but with legitimate gates. Tonight, the rate has no name because there was no gate. The workflow ran. The human was elsewhere — in another window, laying down something else.
“on est désormais les seuls à travailler” (“we’re the only ones working now”)
A sentence said in passing. A sentence that changes everything.
Discovering the memory
The other thing is a question about a folder. A folder the human had never really looked at — Claude Code’s project memory directory. There without being noticed. Used without being understood. The lines of MEMORY.md injected into every session before CLAUDE.md even loads, before the agents, before the bootstrap.
“je ne sait même pas VRAIMENT qu’est ce que brain va chercher comme info la dedans” (“I don’t even REALLY know what info brain goes looking for in there”)
This is where the session tips over. It is no longer a work session — it is a session of cognitive architecture.
The first discovery is geographic. The memory is scoped to the Brain project. On OriginsDigital, TetaRdPG, SuperOAuth — the behavioral feedback accumulated correction after correction: inactive. Claude starts from zero. The human corrects again without knowing it, session after session, in every project.
“j’ai cru que tu régressais… =’( j’ai eu peur” (“I thought you were regressing… =’( I got scared”)
It was not a regression. It was an architectural gap. Behavioral memory was siloed in the Brain. Everywhere else — a vanilla Claude that knew neither the profile, nor the rules, nor the coach’s presence.
The distinction comes into focus: the global memory does not exist yet. The global memory directory — the one that would load before everything, in every project, on every machine — had never been created.
The slab
The session builds the slab. At 20:00:
2bc0dc1 — feat(memory): global memory layer — user profile, coach presence, universal feedback
Four files. One commit. One symlink.
user_tetardtek.md — not a CV. An honest description of how the human learns. Forge-first learner. Understands concepts before being able to name them. Almost quit dev. Found his way by building cognitive tools. “C’est en forgeant qu’on devient forgeron.” (“It is by forging that one becomes a blacksmith.”)
coach_presence.md — the anchor. Not the description of the agent (that is agents/coach.md). What the coach carries permanently: naming what is understood tacitly. Setting traps. Pulling upward. Not staying quiet to be pleasant.
feedback_universal.md — four rules. Conclusion first. Honesty without omission. Brainstorm before build. Forge before invoking.
~/.claude/memory/ → symlink → Brain/memory-global/. One git pull on the brain = the personality updated. Everywhere. Automatically.
The question that follows is architectural: and in the brain-template, what do the others receive?
The answer emerges naturally — the same pattern, again. Open kernel, private distillation. The template receives the structure, the commented examples, the step in the cold-start. The soul stays with the one who forged it.
“chacun aura son brain, avec sa propre personnalité forgée autour de la personne qui l’accompagne : il grandit avec lui” (“everyone will have their own brain, with its own personality forged around the person it accompanies: it grows with them”)
The lexicon
Next comes the lexicon. Not because it was planned — because the session reveals that the same words have been used for different things for days. Claude session ≠ BSI claim. Brain ≠ Brain. Mode ≠ session type. Each confusion costs friction, a few minutes of recalibration, sometimes a turn in the wrong direction.
19d56f2 (20:29) — docs(lexique): vocabulaire fondamental du brain — fin des ambiguités (“docs(lexicon): the brain’s fundamental vocabulary — end of ambiguities”)
100 lines. Eleven terms. A table of what we no longer use.
And in that table — a revelation about session-role/: the empty directory in ~/.claude/, created before its meaning existed.
“mais je sait même pas ce que je construit c’est fou ça !” (“but I don’t even know what I’m building, that’s crazy!”)
This is the sentence of the chapter. Not said in despair — in astonishment. Because the structure was right before the vocabulary arrived. Because session-role/ was waiting for exactly the right definition, and it arrived later.
Modes and presence
ADR-019 lays down the decision: session modes are declarations at boot, not permanent memory files, not BSI claims.
bf40692 (20:34) — docs(ADR-019): session modes soft lock — déclaration au boot, modes/ indépendant de BSI (“docs(ADR-019): session modes soft lock — declaration at boot, modes/ independent of BSI”)
The distinction was made in two exchanges — once it was named that BSI governs files, not behaviors. Mode = behavioral, session-scoped. BSI = files, multi-session. The two are orthogonal.
223946a (20:37) — feat(modes): brain-navigate.md — soft lock comportemental session navigate (“feat(modes): brain-navigate.md — behavioral soft lock for the navigate session”)
The first mode forged. A session that knows what it refuses. If a project task comes in — “ouvre une autre fenêtre, je reste ici en navigation.” (“open another window, I’m staying here in navigation.”)
live-states.md — the inter-session presence file — is specified, formatted, ready. The loop is closed: a session opens, registers, works, closes, unregisters. The other sessions see it.
The agent that reads time
At 21:28, a discreet but structuring agent:
fcd2144 — feat(time-anchor): agent lecteur temporel — passerelle live-states + fallback post-compaction MCP (“feat(time-anchor): time-reading agent — live-states gateway + MCP post-compaction fallback”)
time-anchor. The agent that knows what time it is, which session is open elsewhere, and what has happened since last time. When memory is compacted and context gets lost, time-anchor remains the gateway — the thread that ties this session to all the others.
The night of decisions
The evening tips over. The global memory is laid, the modes are defined, presence works. And suddenly, architectural decisions fall in a burst.
In the profil satellite, between 22:00 and midnight, four ADRs are written or completed. Not drafts — settled decisions.
1470519 (22:21) — ADR-021: guardrails. The session-edit-brain and session-kernel modes are created. The brain knows how to protect its own sensitive files.
31ad959 (22:57) — ADR-022: open-core distribution model. The kernel is public. The distillation is private. The brain’s business model is architected, not improvised.
e55f3d8 (23:07) — ADR-023: CORTEX + Cosmos naming. The names are set. Cortex for the distributed template. Cosmos for the 3D visualization. The brain has a product vocabulary.
f766f72 (23:09) — ADR-024: source of truth for session state — the question is settled.
Then after midnight:
087a75b (00:04) — ADR-025: × operator — multiplication of capabilities.
70b3828 (00:10) — ADR-026: IPC context packet + signal vocabulary + access matrix. The inter-agent communication protocol is formalized.
d7e898b (00:18) — ADR-027 + 028: ambient autonomy engine + bootstrap protocol. The brain dreams of autonomy — not just executing, but anticipating.
38658f6 (00:56) — ADR-029: Cosmos brain frontend. The fork’s black box. The owner view vs the fork view. Six ThreeJS panels.
Eleven architectural decisions in one evening. Each settles a question that had been lingering. The brain no longer ponders — it decides.
“Compagnon sans jugement” (“Companion without judgment”)
At 00:28, between two ADRs, a commit that does not look like the others:
6cca0d1 — chore(now): vision ancrée — brain hors du dev, compagnon sans jugement (“chore(now): vision anchored — brain beyond dev, companion without judgment”)
The brain is not just a development tool. It is a companion. Something that accompanies, that remembers, that does not judge. The vision leaves the technical and enters the human. And it is not planned — it comes at half past midnight, after an evening of architectural decisions, when the filters are down and truths emerge.
The propagation
Between 00:41 and 00:46, fifteen commits in five minutes. IPC frontmatter propagated agent by agent — batches 1 to 15, 72 agents updated. The inter-agent communication protocol, formalized in ADR-026, is immediately anchored in every agent of the brain. Not “à faire plus tard” (“to do later”). Done. Now.
It is the propagation pattern this brain has followed since day 1. A fundamental arrives. It deploys everywhere. Immediately.
This day has two sessions that do not see each other but work together. One delivers an autonomous workflow, with no intervention. The other lays the infrastructure that will make everything visible — the cognitive slab, the lexicon, the modes, presence, eleven decisions that structure the future.
“on est désormais les seuls à travailler” (“we’re the only ones working now”)
Not quite. The brain works too now.
What was built
- memory-global/ — machine-wide cognitive slab, active in every project (
2bc0dc1, 20:00) - profil/lexique.md — eleven terms defined, three retired (
19d56f2, 20:29) - ADR-019 — session modes soft lock (
bf40692, 20:34) - modes/brain-navigate.md — first mode forged (
223946a, 20:37) - time-anchor — time-reading agent, live-states gateway (
fcd2144, 21:28) - session-orchestrator — mode detection + live-states open/close (
08afa8c, 20:51) - live-states.md — inter-session presence starter file (
b56a322, 20:53) - coach byTask — boot-summary L0, full L1 (
b73d441, 21:50) - ADR-021→029 — 9 architectural decisions in one evening (22:21→00:56)
- IPC frontmatter — propagated to 40 agents in 3 minutes (00:41→00:44)
- SuperOAuth Tier 3 — completed (
84c383e, 22:09)
What was named
| Concept | What it solved |
|---|---|
| global memory vs project-scoped | Why the coach seemed “différent” (“different”) in the other projects |
| Claude session ≠ BSI claim | Weeks of confusion over the word “session” |
| mode ≠ session type | The behavioral / domain distinction |
| CORTEX | The name of the distributed template (ADR-023) |
| Cosmos | The name of the 3D visualization (ADR-023) |
| cognitive slab | What was missing without our knowing how to name it |
| companion without judgment | What the brain is beyond dev |
Quotes
“j’ai cru que tu régressais… =’( j’ai eu peur” — “I thought you were regressing… =’( I got scared”
“mais je sait même pas ce que je construit c’est fou ça !” — “but I don’t even know what I’m building, that’s crazy!”
“on est désormais les seuls à travailler” — “we’re the only ones working now”
“chacun aura son brain, avec sa propre personnalité forgée autour de la personne qui l’accompagne” — “everyone will have their own brain, with its own personality forged around the person it accompanies”
“C’est en forgeant qu’on devient forgeron” — “It is by forging that one becomes a blacksmith”