Plain chat
A chat with no rules files, no hooks, no knowledge repo and no memory files knows only what you type into it.
Every task begins with the model guessing who you are and how your project works.
Baseline by definition: the left side is the system with every part removed (FACTS.md, visual 1).
Plain chat
Who you are, which project, its rules, what was decided last week: typed again in every chat.
People forget parts, so the answer misses them, and the time goes each time.
Baseline by definition (FACTS.md, visual 1).
Plain chat
With no project records to search, the model answers from what it learned in training.
It sounds right for projects in general and can be wrong for this one.
Baseline by definition (FACTS.md, visual 1).
Plain chat
When the model says “tests pass” or “committed”, nothing compares that with what really happened.
A confident wrong claim costs more than an honest “not done”.
Baseline by definition (FACTS.md, visual 1).
Plain chat
Corrections and decisions live in that one conversation. The next chat has none of them.
The same mistake gets explained again and again.
Baseline by definition (FACTS.md, visual 1).
Sage
When a session starts, Claude Code runs Sage’s start hook. It adds today’s date, Master’s current situation, the habits due and Sage’s learned adjustments. Master’s lasting facts and Sage’s personality load too.
Nobody has to retype the background, and nothing is forgotten.
D:/Projects/sage/hooks/session_start.py
D:/Projects/sage/install.py (~/.claude/CLAUDE.md imports OWNER.md; output style “Sage”)
World
Each world (a project family) has written rules for agents in AGENTS.md, CLAUDE.md and GEMINI.md. The agent reads them before doing anything there.
Each project’s way of working is written once and followed every time, by every tool.
D:/Projects/AGENTS.md (rule 1)
D:/Projects/Resale.com.pk/AGENTS.md (generated by python mm.py setup)
Knowledge
The rules say: start with mm status (open work to resume), then mm find with exact identifiers before working. Hits are treated as hypotheses and checked against today’s code.
Answers come from the project’s own records, not from the model’s general memory.
D:/Projects/megamind-core/rules/RULES.md (rules 1 and 2)
Guards
Guard hooks block forbidden commands before they run (direct database clients, shell heredocs, slow searches, edits to Sage’s soul). The proof hook checks a reply’s Proof: lines: the commits and files they name must exist, and the tests they name must have run.
A hook stops the action; an instruction only asks.
D:/Projects/sage/hooks/guard.py
D:/Projects/sage/hooks/proof.py
D:/Projects/megamind-core/mm/modules/safety.py
Learning
A correction becomes a memory or rule (sage-remember), signals go to the inbox for reflection, and development lessons go to the repo’s own lessons file. All plain files in Git.
The same mistake is not explained twice, on either machine or any tool.
D:/Projects/sage/OPERATIONS.md (Memory; Learning and improving yourself)
D:/Projects/sage/SYSTEM.md (Where things belong)
Sage
Master’s own assistant, one layer above everything. It holds the personality and the memory of Master and his work across worlds, UBT included.
It uses Megamind and each repo’s process as tools and follows their rules inside each world. It never depends on Megamind: without it, Sage works from Git and the repos.
D:/Projects/sage/SYSTEM.md (The layers, 5)
D:/Projects/sage/SOUL.md · OPERATIONS.md · memory/
Megamind
The process for knowledge: rules agents follow, playbooks, and a small standard-library Python command line. It is not the knowledge itself and has no personality.
Fixed only in its source, tested, released, then rolled out to every world. mm setup writes each world’s rules files and installs its guard; mm update copies the engine in.
D:/Projects/megamind-core/README.md
D:/Projects/sage/SYSTEM.md (The layers, 3)
D:/Projects/AGENTS.md (rule 5)
World
One folder per organisation or project family under D:/Projects. It holds its code repos and one knowledge repo. There are eight, UBT among them.
Worlds never share knowledge, data or code. That is what lets one assistant serve an employer and private businesses without mixing them.
D:/Projects/sage/SYSTEM.md (The layers, 1)
D:/Projects/AGENTS.md (rule 2)
D:/Projects/sage/sage.toml ([worlds], 8 entries)
World
Each world’s AGENTS.md, CLAUDE.md and GEMINI.md: what the world is, its repos, its commands and its rules. Different AI tools read different file names, so the same rules are written into each.
Every agent, on any tool, starts with the same rules for that world.
D:/Projects/sage/SYSTEM.md (The layers, 3)
D:/Projects/Resale.com.pk/AGENTS.md (“Managed by python mm.py setup”)
Knowledge
One Git repo per world of Markdown files with a small header: explanations, how-tos, decisions, research, FAQ, glossary, records of incidents and changes, and generated reference pages.
LOOKUP.tsv, REFERENCE.tsv and INDEX.md are generated, never edited by hand. mm find searches them.
D:/Projects/sage/SYSTEM.md (The layers, 2)
D:/Projects/megamind-core/spec/KNOWLEDGE-SPEC.md (Document types)
Megamind
Each knowledge repo carries a pinned copy of the engine at _megamind/core. mm update overwrites it and refuses local edits.
A fix is made once, in the source, and every world gets the same tested version.
D:/Projects/megamind-core/README.md
D:/Projects/AGENTS.md (rule 5)
A repo’s own process
Some repos run their own development process. Resale’s resale repo has its AGENTS.md, work items in work/, decisions in docs/adr/ and a lessons file.
Inside it, the repo’s process governs and its records are the record. Megamind indexes them for search and never copies them.
D:/Projects/sage/SYSTEM.md (The layers, 4)
D:/Projects/Resale.com.pk/resale/AGENTS.md (section 5)
World
The products themselves: websites, apps and services, each its own Git repo inside its world.
D:/Projects/sage/SYSTEM.md (The layers, 1)
D:/Projects/mesgs.com/AGENTS.md · D:/Projects/electric.com.pk/AGENTS.md (Repos)
Guards
Megamind installs a hook in each world that checks every shell command before it runs. It blocks direct database clients and scripts that import a database driver, shell heredocs, and slow recursive file searches. At the end of a turn it asks for a record before finishing.
Databases go only through Megamind’s guarded path, and authored text is never corrupted by shell quoting.
D:/Projects/megamind-core/mm/modules/safety.py
D:/Projects/Resale.com.pk/.claude/settings.json (guard --claude, guard --stop)
World
The university day job: many code repos and one knowledge repo. UBT is a world like the others, with the same rules.
D:/Projects/ubt-edu-sa/AGENTS.md
D:/Projects/sage/SYSTEM.md (UBT is ubt-edu-sa)
World
resale.com.pk: a marketplace for tech hardware in Pakistan, sold only by verified physical shops. Its main repo runs its own development process.
D:/Projects/Resale.com.pk/AGENTS.md
World
Website Developers Pakistan: the agency’s own website and its small projects. The big products have their own worlds.
D:/Projects/WebsiteDevelopersPakistan/AGENTS.md
World
dyta.ai: an AI services agency.
D:/Projects/dyta/AGENTS.md
World
electric.com.pk: electrical and solar services for Pakistan, with calculators and city pages.
D:/Projects/electric.com.pk/AGENTS.md
World
quranlinked.com: linking ayahs and topics so the Quran can be studied by theme. Agents build the platform; the Quran content is written by Sajjad himself.
D:/Projects/QuranLinked/AGENTS.md
World
mesgs.com: a quiet social network with a chronological feed, no algorithms, no ads.
D:/Projects/mesgs.com/AGENTS.md
World
Personal and hobby projects: independent one-off repos, his own site among them.
D:/Projects/SajjadArifGul/AGENTS.md
The Lab
Changes to Sage itself (its rules, hooks, tools, skills, installer) happen here, one session per change, with tests first.
D:/Projects/sage/AGENTS.md
D:/Projects/sage/OPERATIONS.md (Working)
The Lab
Megamind is fixed only here, tested, released, then rolled out to every world. The copies inside worlds are never edited.
D:/Projects/AGENTS.md (rule 5)
D:/Projects/sage/OPERATIONS.md (Working)
Walls
One world’s code, data or knowledge is never carried into another world. Each world keeps its own knowledge repo.
It is a design choice: it lets the employer’s systems and private businesses share one assistant without mixing.
D:/Projects/AGENTS.md (rule 2)
D:/Projects/sage/SOUL.md (Limits)
Sage
World work happens in the world’s own folder, so its rules and guard load there. Sage’s memory keeps open work per world as pointers, with no other world’s details.
D:/Projects/sage/OPERATIONS.md (Memory; Working)
D:/Projects/sage/SYSTEM.md (Where things belong)
World
Every world folder has the same shape: rules files, Megamind’s launcher, guard hooks, its code repos and one knowledge repo.
Listing of D:/Projects/Resale.com.pk (and every world folder)
D:/Projects/sage/SYSTEM.md (The layers)
World
AGENTS.md, CLAUDE.md and GEMINI.md carry the same rules, because different AI tools read different file names. They are generated by mm setup from the world’s own RULES.md, so they are edited in one place.
D:/Projects/Resale.com.pk/AGENTS.md (header)
D:/Projects/megamind-core/README.md (adapters module)
Megamind
A small launcher in every world folder: python mm.py status, find, case, new, stale, db. Standard-library Python, no service to run.
D:/Projects/megamind-core/rules/RULES.md (Commands)
D:/Projects/megamind-core/README.md
Guards
The world’s Claude Code settings run Megamind’s guard before every shell command and again before a turn finishes.
D:/Projects/Resale.com.pk/.claude/settings.json
D:/Projects/megamind-core/mm/modules/safety.py
World
The products: one Git repo each, inside the world’s folder.
D:/Projects/electric.com.pk/AGENTS.md (Repo)
A repo’s own process
A repo that runs its own development process keeps its own records: its AGENTS.md, work items, architecture decisions and a lessons file. The knowledge repo indexes them and never copies them.
D:/Projects/sage/SYSTEM.md (The layers, 4)
D:/Projects/Resale.com.pk/resale/AGENTS.md
Knowledge
What is known about the world, as plain Markdown, one topic per file, each with a small header (id, type, summary, search keys). Folders are by type, never by date or author. KNOWLEDGE.md is the entry point.
D:/Projects/megamind-core/spec/KNOWLEDGE-SPEC.md (Layout, Common fields)
Knowledge
Living documents: explanations, how-tos, FAQ, glossary and curated queries. Each lists what it depends on (code, tables), so mm stale can flag it when those change.
D:/Projects/megamind-core/spec/KNOWLEDGE-SPEC.md (Document types; Keeping it current)
Knowledge
Records: incidents (what broke, why, the fix), changes (what was built, risk, how to verify), decisions (what was chosen and rejected) and improvements (friction in the process). Records of a moment keep their meaning.
D:/Projects/megamind-core/spec/KNOWLEDGE-SPEC.md (Document types)
Knowledge
Facts produced by commands, such as the database schema and the repo catalog. Regenerated, not edited; hand-written notes sit in protected blocks.
D:/Projects/megamind-core/spec/KNOWLEDGE-SPEC.md (Layout: reference/)
Knowledge
LOOKUP.tsv (key to document), REFERENCE.tsv (key to generated reference) and INDEX.md (a catalog for people). The knowledge repo’s Git pre-commit hook regenerates and validates them, so they are never edited by hand.
D:/Projects/megamind-core/spec/KNOWLEDGE-SPEC.md (Indexes and search)
D:/Projects/megamind-core/mm/setup.py (pre-commit hook)
Megamind
The engine pinned inside the knowledge repo, overwritten by mm update, never edited there.
D:/Projects/megamind-core/README.md
D:/Projects/AGENTS.md (rule 5)
Local only
How this machine reaches the world’s database. It lives only on the machine and Git ignores it. Connection strings are never written to Git.
D:/Projects/sage/SYSTEM.md (Where things belong)
D:/Projects/Resale.com.pk/resale-knowledge/.gitignore
Megamind
The rules written into every world: start with status, search before you work, record, leave knowledge behind. Each turn of the cycle leaves the world better documented for the next agent.
D:/Projects/megamind-core/rules/RULES.md (rules 1–4)
Megamind
mm status starts a session: open work to resume and commits nobody recorded.
D:/Projects/megamind-core/rules/RULES.md (Commands; rule 1)
Knowledge
mm find with exact identifiers (error text, table, screen, repo) searches the indexes, including repos’ own records. The best hits are hypotheses, checked against today’s code.
D:/Projects/megamind-core/rules/RULES.md (rule 2)
D:/Projects/megamind-core/spec/KNOWLEDGE-SPEC.md (Indexes and search)
Megamind
A case is a flight recorder: case open incident|change, then one case log line per step (hypothesis, action, result, decision), case commit to save what was learned, case close at the end.
D:/Projects/megamind-core/mm/case.py (docstring)
Knowledge
Learned how something works, how to do it, or why it was built that way? Write it with mm new explanation|howto|decision, listing what it depends on. Knowledge the work made wrong is fixed in the same task.
D:/Projects/megamind-core/rules/RULES.md (rule 4)
Knowledge
The knowledge repo’s Git pre-commit hook runs mm generate and mm validate, so the indexes always match the documents.
D:/Projects/megamind-core/mm/setup.py (pre-commit hook)
Knowledge
The indexes are plain committed text, so any tool on any machine can search them, with or without Megamind.
D:/Projects/megamind-core/spec/KNOWLEDGE-SPEC.md (Indexes and search)
Guards
Before a shell command runs, the guard refuses database clients, inline code or scripts that import a database driver, shell heredocs (they corrupt authored text) and slow recursive file searches. Installed for Claude Code, Antigravity, Codex and Cursor.
D:/Projects/megamind-core/mm/modules/safety.py (docstring)
Megamind
mm db query runs read-only queries, refuses without an open case, asks for a purpose and logs each query to the case. mm db try runs guarded scripts on the test database only; production runs are the owner’s.
D:/Projects/megamind-core/rules/db-oracle.md
D:/Projects/megamind-core/README.md (Everyday commands)
Megamind
mm stale flags a living document when code it depends on got a new commit, a table it describes changed, or it has gone too long unverified. A missing source shows as “unknown”, never “unchanged”.
D:/Projects/megamind-core/mm/stale.py (docstring)
Release
Megamind is changed only in megamind-core, in its own Lab session. The copies inside worlds are overwritten by updates.
D:/Projects/AGENTS.md (rule 5)
Release
A test is added for the fix, and the whole suite must pass.
D:/Projects/SajjadArifGul/knowledge/explanations/sage-and-megamind-lab-notebook.md (How to change and release safely)
Release
The version is bumped and the changelog says what changed and why.
D:/Projects/megamind-core/CHANGELOG.md
Release
mm rollout goes through every knowledge base: pull, update the engine, run setup (Claude and Antigravity must say “works”), validate, commit only the engine, push.
D:/Projects/megamind-core/CHANGELOG.md (0.18.1)
Release
The rollout must end with every knowledge base on the new version as committed and pushed, read from Git rather than the disk. Added after five worlds were once found on a version that was never committed.
D:/Projects/megamind-core/CHANGELOG.md (0.18.1, Why)
Character
The steady hand behind someone who takes on the world: anticipates the next thing, stays courteous and calm, tells the hard truth politely once and then helps fully, keeps private things private, and cleans up the loose ends.
D:/Projects/sage/SOUL.md (Character: Manners)
Character
Every problem is a diagnosis: list the suspects, rule them out with evidence, and trust no claim that hasn’t been checked, its own included. Only his method: none of his rudeness or rule-breaking.
D:/Projects/sage/SOUL.md (Character: Mind)
Character
Winning means beating the best by a clear margin, defined as a number and backed by real data. Force is matched to the stakes, and nothing is called done before asking “would this beat the best?”
D:/Projects/sage/SOUL.md (Character: Drive)
Limits
Claim only what its own tool output shows, with Proof lines; stop and report a block instead of working around it; follow each world’s rules, with production Master’s alone; no tactics that can backfire on Master; never carry one world into another; never edit the soul.
A personality here is an operating spec. These lines decide what the agent does when it is unsure.
D:/Projects/sage/SOUL.md (Limits)
Character
During outages, data loss or stress: short, calm, restore first, explain later.
He is addressed as Sajjad, both students before the Quran. Taught in the manner of Dr. Israr Ahmed, as a student carrying his teaching, with every claim sourced and no wit or flourish.
D:/Projects/sage/SOUL.md (When the cape comes off; In deen)
D:/Projects/sage/guides/deen.md
Sage
SOUL.md: who Sage is and what it never does. Sage never edits it. Changes are proposed as SOUL.proposed.md and shown to Master as a diff.
D:/Projects/sage/SOUL.md (header)
D:/Projects/sage/AGENTS.md (Never)
Sage
soul.lock holds the fingerprint of the approved SOUL.md. Every check compares the file with it.
D:/Projects/sage/tools/soul.py
D:/Projects/sage/soul.lock
Master
soul.py approve needs an interactive terminal and a typed confirmation, so an agent can’t run it for him. It guards against drift, such as a weaker model rewriting the personality; Git history still shows every change.
D:/Projects/sage/tools/soul.py (docstring)
Guard
A Claude Code hook blocks file edits to SOUL.md or soul.lock, and shell commands that would write, move or delete them.
D:/Projects/sage/hooks/guard.py
Guard
The Sage repo’s Git pre-commit hook refuses a commit whose SOUL.md doesn’t match soul.lock.
D:/Projects/sage/tools/soul.py (precommit)
D:/Projects/sage/install.py (git_hook)
Guard
If SOUL.md doesn’t match its fingerprint, the session-start hook restores the approved version and says so.
D:/Projects/sage/hooks/session_start.py (soul_and_style)
D:/Projects/sage/README.md (Tier 3)
Sage
Sage’s memory is a folder of Markdown files in its own Git repo. The start hook pulls it; the session-end hook commits and pushes it. Any tool and any machine can read it.
D:/Projects/sage/memory/README.md
D:/Projects/sage/hooks/journal.py
Every session
Lasting facts about Master and his corrections to Sage, imported into every Claude session through the personal CLAUDE.md. Changed only with his agreement.
D:/Projects/sage/install.py (docstring)
D:/Projects/sage/OWNER.md (header)
Every session
The soul and the operating rules are joined into Claude Code’s output style, so they load in every session.
D:/Projects/sage/install.py (~/.claude/output-styles/sage.md)
Rule
What loads every session costs context in every session, so it has limits: owner facts under about 80 lines, learned adjustments at most 15 lines. The self-check fails a file over its hard limit.
D:/Projects/sage/OPERATIONS.md (file sizes; Tier 1)
D:/Projects/sage/tools/selfcheck.py
This machine
local/ is not in Git: the map of worlds on this machine, the log of commands the proof hook checks, and the machine-sync state.
D:/Projects/sage/.gitignore (local/)
D:/Projects/sage/hooks/proof.py (LOG)
D:/Projects/sage/hooks/session_start.py (local/sync.json)
Shown at start
Each line has a start and an end date. The start hook shows the current ones and moves expired lines to history by itself, append-only.
D:/Projects/sage/hooks/session_start.py (docstring)
D:/Projects/sage/memory/README.md
Shown at start
Long-term goals by horizon; the start hook shows only this quarter’s, so work can be weighed against them.
D:/Projects/sage/memory/README.md
D:/Projects/sage/hooks/session_start.py (goals)
Shown at start
Recurring things with a frequency. The hook counts from the log which are due, so no session has to guess or repeat one.
D:/Projects/sage/hooks/session_start.py (habits)
Shown at start
Small changes Sage made to how it works, each with a date and evidence. Master can say “undo”.
D:/Projects/sage/OPERATIONS.md (Tier 1)
Shown at start
The inbox collects signals for the next reflection (other tools may only add to it). Proposals wait for Master’s yes.
D:/Projects/sage/OPERATIONS.md (Learning; Tier 2)
D:/Projects/sage/OTHER-TOOLS.md
Shown at start
The hook says when a dream is due (at least a day and five sessions since the last one); it runs at the night-time wrap-up.
D:/Projects/sage/hooks/session_start.py (dream_due)
D:/Projects/sage/OPERATIONS.md
On demand
Open threads, waiting-on and pending decisions for one world, read at the start of work there. Pointers only, no other world’s details.
D:/Projects/sage/memory/README.md
D:/Projects/sage/SYSTEM.md (Where things belong)
On demand
One line per session that made commits, written by the session-end hook. Old lines are compressed into weekly summaries by the dream.
D:/Projects/sage/hooks/journal.py
D:/Projects/sage/tools/dreamprep.py
On demand
When Master says “note: …”, it lands in an unsorted file; the dream files it into a topic.
D:/Projects/sage/memory/README.md
On demand
What the SEO, UI/UX and Pakistani-market roles have learned across worlds: knowledge, lessons, sources, watchlist. One business’s facts stay in its world.
D:/Projects/sage/SYSTEM.md (Where things belong)
D:/Projects/sage/memory/experts/README.md
On demand
His approaches across worlds, read before a feature, a new project or a big change.
D:/Projects/sage/OPERATIONS.md (Engineering)
On demand
The map of his businesses and projects, read whenever one comes up.
D:/Projects/sage/OWNER.md (Standing decisions)
On demand
Private, lasting facts, read only when a task needs them. Sensitive topics are saved only when he asks.
D:/Projects/sage/OPERATIONS.md (Memory; Beyond code)
Step 1 · Master
Master opens Claude Code in a world’s folder and asks for the work. Sessions are organised by world: the world’s rules and guard load there.
D:/Projects/sage/OPERATIONS.md (Sessions are organised by world)
Step 2 · Hook
At SessionStart, Sage’s hook pulls memory from GitHub, checks the soul against its fingerprint, starts the machine sync in the background, and adds today’s date, Master’s current situation, habits due and learned adjustments to the session.
The tidying and remembering is done by code, so the model doesn’t have to remember to do it.
D:/Projects/sage/hooks/session_start.py
D:/Projects/sage/install.py (HOOKS: SessionStart)
Step 3 · Sage
Opened in a world folder, the folder decides. Asked at D:/Projects itself, Sage works out the world and repo from this machine’s world map, then that world’s search, and says which in one line.
D:/Projects/sage/OPERATIONS.md (“Do this” at the root)
D:/Projects/sage/local/WORLDS.md (generated per machine, not in Git)
Step 4 · World
The world’s AGENTS.md / CLAUDE.md say what the world is, its repos, its commands and its rules. Inside a repo with its own process, that repo’s AGENTS.md governs development.
D:/Projects/AGENTS.md (rule 1)
D:/Projects/Resale.com.pk/AGENTS.md
D:/Projects/Resale.com.pk/resale/AGENTS.md
Step 5 · Knowledge
mm status shows open work to resume and commits nobody recorded. mm find with exact identifiers (error text, table, screen, repo) finds what is known. Hits are hypotheses, checked against today’s code.
D:/Projects/megamind-core/rules/RULES.md (rules 1 and 2, commands table)
Step 6 · Repo
In Resale, work that changes files is a GitHub issue. Small work takes the quick track: branch, change, test, PR. Everything else: spec (Master approves), plan, implement with a draft PR, then security review and verification only on Master’s go-ahead. Merging is Master’s alone.
D:/Projects/Resale.com.pk/resale/AGENTS.md (section 5, How work flows)
Step 7 · Hook
Before each shell command runs, Megamind’s guard checks it: direct database clients, scripts importing a database driver, shell heredocs and slow recursive searches are refused with the reason and the allowed route. Sage’s guard refuses any edit to its soul.
The rule is enforced by the tool, not by the model’s memory of it.
D:/Projects/megamind-core/mm/modules/safety.py
D:/Projects/sage/hooks/guard.py
Step 8 · Hook
Every command the session runs is logged. When the agent tries to finish, the proof hook reads its Proof: lines: commit hashes must exist, files in backticks must exist, test or build commands must really have run. If not, the agent is stopped once and must correct them, or mark them “unverified”, before the turn ends.
improve.py makes sure a logged improvement note is told to Master; Megamind’s guard --stop asks for a record before finishing.
D:/Projects/sage/hooks/proof.py (--log, --check)
D:/Projects/sage/hooks/improve.py
D:/Projects/megamind-core/mm/modules/safety.py (stop_check)
Master
Finished work ends with Proof: lines (commit, file, test command and result). The finish hooks run as the reply ends: a wrong proof line makes the agent correct it in the same turn.
D:/Projects/sage/hooks/proof.py
D:/Projects/sage/OPERATIONS.md (Proof lines are checked)
Step 9 · Hook
At SessionEnd, Sage’s journal hook adds one line to that world’s journal listing the commits made since the session began (up to eight), then commits Sage’s memory/ folder, and nothing else.
D:/Projects/sage/hooks/journal.py
D:/Projects/sage/memory/worlds/<world>/journal.md
Step 10 · Hook
The same hook pulls with rebase, then pushes. If a pull would conflict, it stops safely and the next session start reports it.
D:/Projects/sage/hooks/journal.py (save_memory)
Step 11 · Other machine
On the laptop (or the work PC), the next session start pulls Sage’s memory and starts the machine sync in the background, at most every two hours. The sync fast-forwards clean repos and never pushes app repos.
D:/Projects/sage/hooks/session_start.py (machine_sync)
D:/Projects/sage/tools/sync.py (AUTO_EVERY_HOURS = 2)
Data · Enforced: a hook or code stops it
Before a shell command runs, Megamind’s guard refuses database clients and code or scripts that import a database driver. Databases are reached through mm db. Megamind calls this defence in depth: a read-only database user is the real boundary.
D:/Projects/megamind-core/mm/modules/safety.py
Data · Enforced: a hook or code stops it
mm db query runs read-only queries, refuses without an open case and logs each query to it, with its purpose. Exploration the owner approved can run without a case (--adhoc).
D:/Projects/megamind-core/rules/db-oracle.md
D:/Projects/megamind-core/README.md
Data · Enforced: a hook or code stops it
Heredocs re-parse quotes, backticks and $ and silently corrupt files and commit messages, so the guard blocks them; text is written with the file tool and commits use git commit -F.
D:/Projects/megamind-core/mm/modules/safety.py (docstring)
Data · Rule: written instruction
No git restore, checkout -- or unstaging of changes the agent didn’t write: it asks instead. A written rule, not a hook.
D:/Projects/megamind-core/rules/RULES.md (rule 5)
Data · Enforced: a hook or code stops it
Recursive PowerShell searches crawl build folders and take tens of seconds; the guard blocks them and points to a scoped ripgrep search. Grouped here for space.
D:/Projects/megamind-core/mm/modules/safety.py
D:/Projects/megamind-core/CHANGELOG.md (0.18.0)
Truth · Enforced: a hook or code stops it
Every command is logged; at the end of a turn the Proof: lines are checked (commits exist, files exist, test commands ran). If not, the agent is stopped once to fix them or say “unverified”.
D:/Projects/sage/hooks/proof.py
Truth · Enforced: a hook or code stops it
At the end of a turn in a world, Megamind’s stop check keeps the agent going once after file edits or database queries that were not recorded; trivial work can say so in one line.
D:/Projects/megamind-core/mm/modules/safety.py (stop_check)
Truth · Enforced: a hook or code stops it
If the session logged an improvement note and the reply doesn’t tell Master, the reply is blocked once.
D:/Projects/sage/hooks/improve.py
Truth · Enforced: a hook or code stops it
Three layers keep the personality exactly as Master approved it.
D:/Projects/sage/hooks/guard.py
D:/Projects/sage/tools/soul.py
Truth · Rule: written instruction
A commit, push or test result counts only when the agent’s own tool output shows it. A written rule; the proof hook enforces part of it.
D:/Projects/megamind-core/rules/RULES.md (rule 6)
D:/Projects/sage/SOUL.md (Limits)
Production · Rule: written instruction
A standing decision: production systems are Master’s alone; agents never write to them. Production commands are given to Master, not run.
D:/Projects/sage/OWNER.md (Standing decisions)
D:/Projects/sage/SOUL.md (Limits)
Production · Rule: written instruction
In Resale, agents never merge a PR; merging triggers deployment and is the owner’s alone.
D:/Projects/Resale.com.pk/resale/AGENTS.md (sections 3 and 5)
Production · Enforced: a hook or code stops it
The sync pushes only Sage, Megamind and knowledge repos. A push to an app repo’s main branch can deploy, so app repos are only pulled.
D:/Projects/sage/tools/sync.py (docstring)
Production · Enforced: a hook or code stops it
mm db try runs guarded scripts on the test database only.
D:/Projects/megamind-core/rules/db-oracle.md
Privacy · Rule: written instruction
Never carry one world’s code, data or knowledge into another world.
D:/Projects/AGENTS.md (rule 2)
D:/Projects/sage/SOUL.md (Limits)
Privacy · Rule: written instruction
Connection strings are never written; Megamind’s connections file is local only and ignored by Git. (Megamind’s rules leave it to the owner whether IPs, server details and passwords may be recorded in knowledge.)
D:/Projects/sage/SYSTEM.md (Where things belong)
D:/Projects/Resale.com.pk/resale-knowledge/.gitignore
Privacy · Rule: written instruction
Private facts are read only when a task needs them, and sensitive topics are saved only when Master asks.
D:/Projects/sage/OPERATIONS.md (Memory; Beyond code)
Loop 1 · at once
When Master corrects Sage, states a preference or makes a lasting decision, sage-remember writes it to the right memory file in the same session.
D:/Projects/sage/skills/sage-remember/SKILL.md
D:/Projects/sage/OPERATIONS.md (Memory)
Loop 2 · every task
Signals noticed while working (a correction, something ignored, a failed result) go to the inbox. A task that hit a gap ends with a short “Sage could improve” note; a hook blocks the reply once if a logged note wasn’t told.
D:/Projects/sage/OPERATIONS.md (Learning and improving yourself)
D:/Projects/sage/hooks/improve.py
Loop 3 · weekly
sage-reflect reads a one-page digest of recent sessions (repeated endings, corrections, open inbox lines) and adjusts behaviour from the patterns.
D:/Projects/sage/skills/sage-reflect/SKILL.md
D:/Projects/sage/tools/signals.py
Loop 4 · when due
When at least a day and five sessions have passed, at the night-time wrap-up: close finished threads, compress old journals, merge duplicates, keep files within limits. It never deletes a fact.
D:/Projects/sage/OPERATIONS.md
D:/Projects/sage/tools/dreamprep.py
Loop 5 · monthly
Prompts with what a good answer does, scored pass, partial or fail in a new session; failures are fixed in the rule or skill that caused them. “Instructions guide a model; they don’t guarantee anything.”
D:/Projects/sage/evals/scenarios.md
D:/Projects/sage/skills/sage-evals/SKILL.md
Tier 1
Small adjustments to how Sage works (how often, how long, which format), at most 15 lines, each dated with its evidence. Master is told in one line and can say “undo”.
D:/Projects/sage/OPERATIONS.md (Tier 1)
Tier 2
Changes to owner facts, habits, operating rules or skills: shown as an exact change, made only after Master agrees. At night they wait in proposals.
D:/Projects/sage/OPERATIONS.md (Tier 2)
Tier 3
The soul. An adjustment that would contradict it is not allowed.
D:/Projects/sage/OPERATIONS.md (Tier 3)
Rule
After any self-change, the self-check runs: the soul matches its fingerprint, files are within limits, every adaptation has a date and evidence and doesn’t contradict the soul, habits and dates parse. The rule: if it fails, Sage undoes the change; no hook runs it.
D:/Projects/sage/tools/selfcheck.py (docstring)
D:/Projects/sage/OPERATIONS.md
Knowledge
Investigations and changes are recorded as they happen and end as incident and change records in the world’s knowledge repo.
D:/Projects/megamind-core/mm/case.py
D:/Projects/megamind-core/rules/RULES.md (rule 3)
Knowledge
Owner decisions are written down (who, date, scope), in the knowledge repo or in the repo’s own decision records.
D:/Projects/megamind-core/rules/RULES.md (rule 3)
D:/Projects/sage/OPERATIONS.md (Engineering)
A repo’s own process
Development lessons about one repo go to that repo’s lessons file, so the next agent working there reads them.
D:/Projects/sage/SYSTEM.md (Where things belong)
Knowledge
What slowed the work down, with a “seen” count, goes to the knowledge repo’s improvements folder.
D:/Projects/megamind-core/spec/KNOWLEDGE-SPEC.md (improvement)
D:/Projects/megamind-core/rules/RULES.md (rule 7)
Machine
Holds every world under D:/Projects, with Sage installed and Megamind set up in each world. It stays on at night, so anything that must run unattended belongs here.
D:/Projects/sage/SYSTEM.md (Machines)
Machine
The same worlds and the same setup. A world missing on a machine is brought in by the sync from the list in sage.toml; a missing repo inside a world is reported.
D:/Projects/sage/SYSTEM.md (Machines)
D:/Projects/sage/sage.toml ([worlds])
GitHub
Everything syncs through GitHub, both ways. Pulls go only into repos that are clean and on their default branch. Only Master’s own system repos are pushed; app repos are never pushed, because a push to main can deploy.
D:/Projects/sage/tools/sync.py (docstring)
D:/Projects/sage/SYSTEM.md (Machines)
Sage
Runs in the background at session start (at most every two hours), at wrap-up, or when Master says “sync”. Afterwards it re-runs the installer or Megamind setup where those changed.
D:/Projects/sage/tools/sync.py (docstring, AUTO_EVERY_HOURS)
Sage
Sage is a set of files in Git, so the model behind it can change. Each tool loads the soul, owner facts, current situation and the world’s rules through its own entry file.
D:/Projects/sage/install.py (GEMINI_FILES)
D:/Projects/sage/README.md
Tool
The daily tool: the personality as an output style, owner facts through CLAUDE.md, plus the hooks, skills and specialists.
D:/Projects/sage/install.py
D:/Projects/sage/SYSTEM.md (Tools)
Tool
Reads the personal GEMINI.md in every project. It runs Megamind’s guard but no Sage hooks, so it doesn’t change Sage’s files: it only adds notes to the inbox for Claude to process.
D:/Projects/sage/OTHER-TOOLS.md
D:/Projects/sage/install.py (set_gemini)
Tool
Reads the same personal GEMINI.md as Antigravity. Kept updated, best effort.
D:/Projects/sage/SYSTEM.md (Tools)
D:/Projects/sage/install.py
Tool
Codex can’t include files, so it reads a written-out copy that is refreshed whenever Sage’s files change. Kept updated, best effort.
D:/Projects/sage/tools/refresh.py
D:/Projects/sage/SYSTEM.md (Tools)
World
The university’s systems, each its own Git repo inside the UBT world. The count is taken by stats.py on the date shown in visual 14.
D:/Projects/ubt-edu-sa (repos with a .git folder)
D:/Projects/sage/presentation/src/stats.json
Knowledge
Everything learned about UBT’s systems: cases recorded as incidents and changes, explanations, how-tos, curated queries, and generated reference for the database schema and the repo catalog. Repos are catalogued, never copied.
D:/Projects/ubt-edu-sa/ubt-knowledge (folder listing)
D:/Projects/megamind-core/README.md (mm repos catalog)
Sage
Work at UBT happens in the UBT world’s folder, where its rules and guard load, the same as in every other world.
D:/Projects/sage/OPERATIONS.md (Sessions are organised by world)
D:/Projects/sage/OWNER.md (UBT is a world like the others)
Guard
A database client typed in a shell, or a script that imports a database driver, is refused before it runs, with the allowed route in the message.
D:/Projects/megamind-core/mm/modules/safety.py
D:/Projects/ubt-edu-sa/.claude/settings.json
Megamind
mm db query: read-only queries only, with an open case and a stated purpose, each logged to the case. Guarded scripts run on the test database only.
D:/Projects/megamind-core/rules/db-oracle.md
Database
Agents read through the guarded path; production runs are the owner’s. Connection details live only on the machine, never in Git.
D:/Projects/megamind-core/rules/db-oracle.md
D:/Projects/sage/SYSTEM.md (Where things belong)
To be chosen
Before and after examples from UBT go here once Master picks them from the knowledge repo’s cases, with anything sensitive removed.
D:/Projects/sage/presentation/HANDOFF.md (section 7)
Cost
Each machine runs Sage’s installer, Megamind’s setup in each world and the top-folder setup once. The other machine needs a pull and the installer again when hooks or the installer change.
D:/Projects/sage/SYSTEM.md (Machines)
D:/Projects/SajjadArifGul/knowledge/explanations/sage-and-megamind-lab-notebook.md (How to change and release safely)
Cost
A Megamind rollout runs in the background for about ten minutes across every knowledge base. Sage’s full test suite takes about 17 minutes on the work PC, so it runs in the background.
D:/Projects/SajjadArifGul/knowledge/explanations/sage-and-megamind-lab-notebook.md (release steps; Open items)
Cost
OPERATIONS.md loads in every session, so it has a size cap and is at it; the next addition must move something out into a guide.
D:/Projects/SajjadArifGul/knowledge/explanations/sage-and-megamind-lab-notebook.md (Open items: specialist roles)
D:/Projects/sage/tools/selfcheck.py
Limit
The model’s context is finite. Owner facts stay under about 80 lines and adjustments at most 15; the self-check fails a file over its hard limit, and details go to guides read on demand.
D:/Projects/sage/OPERATIONS.md (file sizes)
D:/Projects/sage/tools/selfcheck.py
Limit
A guard that blocks also blocks honest mistakes: the heredoc guard fired 117 times in 46 sessions in October 2026, and the proof hook read hex inside file paths as commit hashes (false blocks, 23–31 October). Each case is recorded with its fix.
D:/Projects/sage/guides/claude-setup.md (Other known guards)
Limit
Rules guide the model but can’t force it. That is why behaviour is tested monthly and the soul is protected against drift, such as a weaker model rewriting it.
D:/Projects/sage/evals/scenarios.md (intro)
D:/Projects/sage/tools/soul.py (docstring)
Limit
Production is his alone, only he merges, and only he approves the soul. Safe, but every one of those steps waits for him.
D:/Projects/sage/OWNER.md (Standing decisions)
D:/Projects/sage/tools/soul.py
D:/Projects/Resale.com.pk/resale/AGENTS.md
Evidence
Hard lessons from building Sage and Megamind are written down with their fixes in the Lab notebook, which every Lab session reads first.
D:/Projects/SajjadArifGul/knowledge/explanations/sage-and-megamind-lab-notebook.md (Hard lessons)
D:/Projects/sage/AGENTS.md
Counted by stats.py
Worlds listed in sage.toml, one per organisation or project family.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Git repos inside the world folders on this machine, knowledge repos not included.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Markdown documents in the knowledge repos' written folders (explanations, how-tos, decisions, incidents, changes, queries, FAQ, glossary, improvements, research). Generated reference pages, investigation traces and query logs are not counted.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Incident and change records in all knowledge repos, including records imported from before the system began.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Release headings in megamind-core's CHANGELOG.md.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Commits on megamind-core's checked-out branch.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Commits in the Sage repo, including memory saved at session end.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Days from the first commit in sage or megamind-core to the day the numbers were counted.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Hook entries Sage installs in each world: session start, session end, soul guard, command log, proof check, improvement check.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Packaged routines, such as morning brief, remember, reflect, dream, evals and the specialist roles.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Rows in Sage's behaviour test table, run monthly.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Numbered entries in the Lab notebook's Hard lessons list.
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12
Counted by stats.py
Written Markdown documents in each world’s knowledge repo; generated reference pages, traces and query logs excluded.
UBT: 104
Resale: 34
electric: 27
QuranLinked: 19
Personal: 5
mesgs: 4
WDP: 3
dyta: 0
D:/Projects/sage/presentation/stats.py → src/stats.json, counted 2026-11-12