Gemini)#

All three AI agents load this file at session start — it is the single entry point into the knowledge base.

Who the user is#

Name: Peti (user@example.com)
Language: Hungarian primarily, English technical terms are OK
Server: /root headless Linux, VSCode extension + SSH + occasional VNC
Style: terse responses, explanation only when needed. Details in 05-Memory/User

Vault structure (Johnny-Decimal folder-prefix + Karpathy LLM-Wiki pattern)#

/root/obsidian-vault/
├── README.md                     human entry point
├── AGENTS.md                     this file (stays at root — symlink-compatible)
├── 00-Meta/                      vault rules
│   ├── README.md
│   ├── Tag-taxonomy.md           MANDATORY tag conventions (#env/prod, #type/host, etc.)
│   ├── Frontmatter-schema.md     MANDATORY YAML schema per type
│   ├── Glossary.md               slug + abbreviation resolver (KGC, MFL, MAPESZ, BMAD…)
│   └── templates/                Daily / Session / Project templates
├── 01-Daily/                     daily logs (YYYY-MM-DD.md, autogen)
│   └── Index.md
├── 02-Projects/                  READ THIS FIRST
│   ├── Index.md                  every project + status in 5 groups
│   └── <project>.md
├── 03-Hosts/                     VPSes + shared hosting
│   ├── Index.md
│   └── <host>.md
├── 04-Tasks/
│   ├── Backlog.md                Obsidian Tasks plugin
│   └── Dashboard.md              filtered query views
├── 05-Memory/                    persistent context
│   ├── README.md
│   ├── User.md                   user profile, preferences
│   ├── Infrastructure.md         server, VNC, ports, KGC Postgres
│   ├── Skill-map.md              243+ agent-skills grouped
│   ├── Agents-skill-suite.md     skills available on myforge-dashboard
│   └── Dashboard-access.md       Tailscale-only access protocol
├── 06-Audits/                    snapshot reports
│   ├── Index.md
│   └── System_Health.md          weekly auto-gen vault integrity
├── 07-Decisions/                 ADR-style decision log
├── 08-Sessions/                  /11.11-session logs (test → _archive/)
├── 10-raw/                       Karpathy: raw input (immutable)
│   └── Index.md
└── 11-wiki/                      Karpathy: distilled knowledge (evergreen)
    └── Index.md

What to do AT SESSION START#

MANDATORY workflow (after /11.11-uj-session the agent automatically):

Detect the project from the session name based on the project-detection table
Aggressive pre-load — read according to wiki/Auto-context-loading:
Project file + last 5 sessions + every relevant ADR + relevant slice of Memory + #project/<slug> Backlog entries + Host info + today's/yesterday's Daily
Target: ~15-20K token context
Write a ## Pre-loaded context section into the session file — list what you read, with a 1-2 sentence summary of each
Only then wait for the user's first question — with full context

If an unknown abbreviation / slug comes up → meta/Glossary. If the project is ambiguous → ask back.

If no project can be detected (e.g. "wellbeing", "general thinking") — load only base context: ../02-Projects/Index, urgent items from ../04-Tasks/Backlog, today's + yesterday's Daily.

When to WRITE here#

Whenever some long-term useful info comes up:

Event	Where to write
New project starts	New `02-Projects/<name>.md` + row in the matching group table in ../02-Projects/Index
Existing project status changed	Update `02-Projects/<name>.md` "Current state" + `updated:`
Learned a user preference	05-Memory/User or new `05-Memory/Feedback-<topic>.md`
Server/infra knowledge (port, host, service)	05-Memory/Infrastructure
Major architecture decision	New `07-Decisions/YYYY-MM-DD <topic>.md`
New TODO	../04-Tasks/Backlog (red urgent / yellow watch / green hygiene)
Task done	../04-Tasks/Backlog — Completed section with date
New concept / playbook (evergreen)	`11-wiki/<topic>.md` — in your own words
New article / transcript (immutable)	`10-raw/YYYY-MM-DD — <source>.md`

Formatting conventions#

Date: YYYY-MM-DD ISO; convert relative dates from the user (e.g. "Thursday") to ISO
Wikilink: [[../02-Projects/teszt-eu]] for internal references (folder-prefix mandatory)
Frontmatter: every larger doc must have name:, type:, created:, updated: (see meta/Frontmatter-schema)
Tags: per meta/Tag-taxonomy
Callout: > [!info], > [!warning], > [!todo], > [!success]
Code/path: in backticks

Crystallization workflow (MANDATORY at session close)#

MANDATORY workflow (after /11.11-zar-session the agent automatically):

Summary + Learnings + Next written into 08-Sessions/<slug>.md based on chat history
(optional, if VAULT_CRYSTALLIZE_AUTO=1) 11.11crystallize <slug> --scorer claude-code --with-context — automatic G-Eval scoring in shadow mode. If a pending request is created, spawn a general-purpose Agent with the G-Eval prompt + bullets + kodb_context; have it write response.json; then re-run 11.11crystallize with the same flag to finalize.
Routing — for every Learning bullet apply the decision tree (architecture-level ADR / vault rule / wiki concept / glossary / infra / skill / user-pref / dashboard / project / task / ask). If a G-Eval output exists, use it as signal (high-confidence Pass → auto-prop candidate; Fail → discard; batch-preview → present to user)

Batch preview for the user — all proposals together, numbered:

N learnings to propagate — I suggest these:
[1] "<quote>" → <target> + preview
[2] "<quote>" → <target> + preview
...
OK like this? (yes / "1-3 OK, 4 rather X" / "skip 2" / "stop")

After user confirmation propagate the knowledge
## Propagation log section — timestamped, record what was propagated where

The session file remains a raw-like reference — the distillate goes into the indexed layers. This is Karpathy crystallization at its minimum.

Detailed rules: wiki/Crystallization-protocol

SV B-1 pipeline (LIVE since 2026-05-16)#

Layer	Command	Function
0 migration	`migrate-hash-refactor-2026-05-19.py`	KO-DB fact-hash refactor: hash by `(s,p,o)` only, `fact_provenance` 1:N side-table (2026-05-19, 190 ms migration, unlocks Bayesian #21)
1 ingest	`vault-ko-ingest --file <path>`	Triplet-extraction via subagent, 2-phase pending pattern
2 score	`11.11crystallize <slug> --scorer claude-code --with-context`	G-Eval Learning-bullet scoring + KO-DB context-inject
3 query	`vault-ko-query <pattern>`	Substring + filter + JSON + `--stats` + `--conflicts` + `--top-k` (cross-source rank) + `--semantic` (Memgraph bridge)
3 report	`vault-ko-report [--last\\|--session <slug>\\|--days N]`	User-facing audit-log summary
4 apply	`VAULT_CRYSTALLIZE_APPLY=1 VAULT_CRYSTALLIZE_REAL=1 11.11crystallize ... --apply`	REAL mode (sandbox-branch + 4-layer safety-gate + atomic-write + auto-commit)
5 monitor	`vault-crystallize-monitor [--weeks N] [--json]`	Auto-rate / revert-rate / threshold-ramp recommendation
5 conflicts	`vault-ko-conflicts-audit`	Weekly cross-source contradiction audit, predicate-aware heat-classifier
5 revert	`crystallize-revert <bullet-hash>`	Auto-apply rollback (with audit-event)

Threshold-config: ~/.vault-config/crystallize-threshold.txt (hot-reloadable). Shadow=1.0 (default, no auto-prop), Conservative=0.95, Aggressive=0.85. Ramp protocol: wiki/crystallize-threshold-ramp.

Env-vars (opt-in): - VAULT_CRYSTALLIZE_AUTO=1 — 11.11stop runs the scoring automatically - VAULT_CRYSTALLIZE_SCORER=claude-code — subagent-fanout scorer ($0 cost) vs mock (rule-based) vs anthropic (API-key required) - VAULT_CRYSTALLIZE_APPLY=1 — enables the --apply flag (Layer 1 ENV gate) - VAULT_CRYSTALLIZE_REAL=1 — REAL mode (write + commit). Default is skeleton-only ("would-have-applied" audit log) - VAULT_CRYSTALLIZE_ALLOW_MAIN=1 — --apply REAL run on main (DEFAULT: only on crystallize-sandbox-* branch). Dangerous.

SV B-2 pipeline (LIVE since 2026-05-13, B-1↔B-2 bridge 2026-05-17)#

Layer	Command	Function
1 embed	`vault-embed --backfill <dir>`	bge-m3 multilingual embed → Memgraph
2 search	`vault-search "<query>" [--top-k N] [--json]`	semantic cosine-search (Memgraph in-Python)
3 bridge	`vault-ko-query "<q>" --top-k N --semantic`	semantic → KO-DB top-K bridge (LIKE-fallback if Memgraph down)

Virtual-context (Letta/MemGPT) — PREVIEW (opt-in, since 2026-05-25 Day 0)#

Alongside the classic aggressive ~17K-token pre-load there is an alternative: core-memory + on-demand page-in. Sprint plan: decisions/2026-05-25 vault-core-memory MemGPT integration sprint plan. Day 0 LANDED, integration between W1-5.

Layer	Command	Function
1 core	`vault-core-memory show`	~2 KB always-loaded core (6 blocks: user_profile / active_project / open_tasks / glossary / infra_pins / recent_decisions)
2 page-in	`vault-core-memory page-in "<query>" [--top-k N] [--max-chars N]`	Real Memgraph chunk-text retrieval, agent-paste-ready markdown
3 simulate	`vault-core-memory simulate "<query>"`	Token-budget A/B (classic vs virtual) — does NOT mutate context
4 update	`vault-core-memory update <block> "<text>"`	Atomic block mutation

Env-var (opt-in, default OFF in Week 1-4): - VAULT_CORE_MEMORY_AUTO=1 — 11.11focus + /11.11-uj-session auto-update + lean pre-loaded context (activates from W1-2)

When to use page-in (agent convention from W2): - If the question CANNOT be answered from core-memory (query-overlap with the 6 blocks alone is not a hit) → page-in "<query>" - The archival page-in result gives ~2-3 KB of chunk-text — instead of the classic 15-20K aggressive pre-load - If >20 page-faults are needed within one session → flag that the core-memory is mis-sized (review required)

Status: Day 0 (2026-05-25) — page-in LANDED, integration not active (env-var OFF). See wiki/vault-core-memory-integration-roadmap.

Net-learning (since 2026-05-17)#

vault-net-ingest --url <URL> — firecrawl scrape → 10-raw/external/
vault-net-ingest --repo owner/name [--paths "globs"] — gh clone → key markdown → 10-raw/external/<owner>_<name>/
KO-DB extraction follows the subagent-fanout-pattern (2-phase pending) — triplets are uploaded after claude_code_scorer_load_response

Per-chat session isolation (since 2026-05-17)#

Claude Code: $CLAUDE_CODE_SESSION_ID env-var (UUID)
Codex companion: $CODEX_COMPANION_SESSION_ID (= Claude UUID)
Codex standalone: vault-detect-chat-id auto-detect (from rollout filename)
Gemini hook: ~/.gemini/.current-session-id (written by SessionStart hook)
Manual override: CODEX_SESSION_ID or GEMINI_SESSION_ID env-var
The 6 11.11* scripts automatically extract the CHAT_ID from the chain, per-chat pointer file: .active-session-$CHAT_ID
Matrix doc: wiki/cli-session-id-env-var-matrix

What NOT to store here#

Secrets (passwords, tokens) — those stay in .env files. Exception: if the user explicitly requests it
Generic info that can be extracted from git log or code
Momentary session state that will be irrelevant in 1 day

Tool-specific notes#

Claude Code: also sees a copy of this file as ~/.claude/CLAUDE.md. The legacy auto-memory (~/.claude/projects/-root/memory/) also lives off this vault via symlinks
Codex CLI: sees it as ~/.codex/AGENTS.md. The ~/.codex/skills/ and ~/.claude/skills/ are updated together via shared symlinks
Gemini CLI: sees it as ~/.gemini/GEMINI.md. Same skill set, same vault

Session orchestration (`11.11*` commands)#

Detailed description: wiki/11.11-session-protokoll

[!info] Slash vs shell The shell CLI (/usr/local/bin/11.11*) names are unchanged: 11.11start, 11.11stop, 11.11focus, 11.11note, 11.11ls, 11.11. The Claude Code slash commands received longer Hungarian names from 2026-05-17 (due to the popup UI — the VSCode extension only shows the command name, NOT the description):

Slash command (Claude Code)	Shell CLI	Function
`/11.11-uj-session "project-task"`	`11.11start "..."`	New session file + focus. Without arg → interactive picker
`/11.11-lista`	`11.11ls`	Open sessions + last 5 closed
`/11.11-focus <slug>`	`11.11focus <slug>`	Switch focus to another open session. Without arg → interactive picker
`/11.11-jegyzet "..."`	`11.11note "..."`	Timestamped note into the focused session's `## Events`
`/11.11-jegyzet @<slug> "..."`	`11.11note @<slug> "..."`	Note targeted at a specific session (substring match)
`/11.11-zar-session [slug]`	`11.11stop [slug]`	Agent: Summary + Learnings + Next; Script: commit + push + close. Without arg + 2+ open → picker
`/11.11-egeszseg`	`11.11`	Health check: vault, symlinks, skills, services

Scripts: /usr/local/bin/11.11* + /11.11* symlinks. Agent env var: AGENT=claude|codex|gemini goes into the commit message.

Vault health#

Live snapshot: audits/System_Health — weekly cron vault-cleanup regenerates (Sunday 04:00)
Auto-save: every 10 minutes cron /usr/local/bin/vault-autosave → commit + push to GitHub
Manual check: /11.11-egeszseg (slash) or 11.11 (shell) — symlinks, skills, services

How this file gets updated#

If a new convention emerges, edit it directly — the change is live for all three agents immediately (symlink).

README — human entry point
meta/README — vault rules
wiki/Karpathy-LLM-Wiki-pattern — the underlying pattern
wiki/Johnny-Decimal-prefix — why the 00-, 01-, … prefix
wiki/11.11-session-protokoll — session organization in depth
../02-Projects/Index — project dashboard