ARIA is a symbolic, adaptive AI · version 1.3.1: a declarative catalog of 14 skills, an intent router with proof, a rule engine, a disciplined memory and a security self-audit · no LLM, no machine learning model, no cloud. It learns exclusively from the conversations you have with it and runs 100% on your machine.
🇫🇷 Une version française de ce README est disponible : README.fr.md
- Declarative skill catalog · every skill is a
skills/<name>/SKILL.mdfolder with frontmatter (name · description · risk · command · triggers): declaring it is enough to make it routable, the catalog validates consistency and lists invalid declarations. Shipped (14): · utilities: date/time, current-month calendar (today marked), unit converter (km/mi, m/ft, kg/lb, °C/°F, Mo/Go), world clock (zoneinfo, no network); · project & docs: local task list (add / list / complete), change journal draft built from local git history, session report; · review & analysis: charter inspector (em-dashes, artifacts, credential literals · positions only, never the content), text statistics, CSV preview; · search & math: DuckDuckGo web search, calculator (safe AST evaluator); · local randomness: value generator (secretsmodule, displayed · NEVER stored), draws (dice, coin flip, 1 to 100). - Intent router with proof · skills, logic and learned rules are all
candidate routes with a confidence score; only the winner runs (lazy)
and the choice is traceable
(
[route: … · risk: … · confidence: … · source: …]). - IF/THEN rule learning · answers are matched against learned rules;
every unknown input is a teaching opportunity (
Réponse correcte ?→ new rule written to disk). Feedback only applies to learned rules: no misleading duplicates. - Memory discipline · secrets are never stored (filter on BOTH fact and rule writes, explicit refusal), hygiene bounds (200 facts · 500 rules), every erasure request ("oublie …") is honored without hidden copies, session compaction with topic archives.
- Output masking · any sensitive value that would be returned is masked line by line (defense in depth).
- Security self-audit ·
python main.py --audit: six offline phases (surface survey → guided hunt → reproducible validation →data/findings.json→ independent re-read → neutral report) chasing secrets indata/and dangerous calls. - Persistent memory · rules, facts, experiences and context are plain
JSON files in
data/, readable and editable by hand. - Fuzzy understanding ·
difflib.get_close_matchesmaps typos and variants to known rules ("qu'elle heure" → "quelle heure est-il"). - Automatic calculations · detects expressions anywhere in the input
("1550 + 12500" →
Le résultat est : 14050) through a safe AST evaluator (no raweval, bounded exponents · see changelog). - Fact-based reasoning · statements ("je m'appelle Martin") are stored as facts and later recalled ("quel est mon nom" → "Tu t'appelles martin.").
- Web search · "recherche python" triggers a real DuckDuckGo Instant
Answer query over
requests(the only network dependency). - Reflection · a short-term memory of the last exchanges gives ARIA self-observations: it flags unknown answers, notices repeated topics, asks follow-up questions.
- Special commands · time/date, open browser, session report, web
search (skill mapping
CMD:*handled bycore/skills.py). - Two interfaces · console (works without tkinter) and Tkinter GUI.
- Diagnostic modes ·
--debug(decision trace, fuzzy matches, facts),--stats(memory state + catalog summary),--audit(security self-audit).
Real Tkinter window, captured during a genuine session · the "1550 + 12500" calculation and the remembered name are live answers:
git clone https://github.com/Mr-Jacks0n/IA.git
cd IA
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt # requests is the only dependency
python main.py # console mode
python main.py --gui # graphical mode (needs python3-tk)
python main.py --debug # console + decision logs
python main.py --stats # memory state + catalog summary
python main.py --audit # security self-audit (report data/findings.json)
python utils/clean_data.py # lowercase-normalize saved rulesConsole works on servers with no tkinter installed; --gui prints
an actionable install hint (sudo apt install python3-tk) if Tk is absent.
.
├── main.py # entry point: console | --gui | --debug | --audit | --stats
├── config.py # DEBUG_MODE + debug_print helper
├── skills/ # 14 declarative skills (SKILL.md)
│ ├── annonce-date/ · rapport-session/ · calculatrice-sure/ · recherche-web/
│ ├── gestion-taches/ · journal-changements/ · inspecteur-charte/
│ ├── analyse-texte/ · apercu-csv/ · convertisseur-unites/
│ └── fuseaux-horaires/ · calendrier-mois/ · generateur-jetons/ · tirage-hasard/
├── core/
│ ├── decision_engine.py # pipeline: winning route → execution → verification
│ ├── router.py # candidate routes, scores, routing proof
│ ├── catalog.py # frontmatter parser + validation + secret guard
│ ├── rules.py # DEFAULT_RULES (seed rules)
│ ├── memory.py # disciplined JSON memory (rules/facts/context)
│ ├── reflection.py # short-term memory + self-observations
│ ├── selfaudit.py # security self-audit (6 phases, offline)
│ └── skills.py # date, calculator (AST-safe), web search, browser, report
├── interface/
│ ├── console.py # REPL + oui/non feedback loop (clean EOF)
│ └── gui.py # Tkinter chat window
├── learning/feedback.py # learn(rule) → Memory.add_rule (MemoryRefused on secrets)
├── utils/clean_data.py # normalize saved rule keys
├── data/*.json # persistent state (defaults shipped)
├── docs/ # banner + real GUI screenshot
└── .github/ # CI + issue/PR templates
Every push runs a 3-version CI (Python 3.10 → 3.12) that:
- lints the codebase (flake8
E9/F63/F7/F82, blocking), - runs a scripted console session (greet → teach confirmation → calculation → name memorization → recall → date command → clean exit) and asserts the exact expected answers,
- runs the security self-audit (
python main.py --audit) and validatesdata/findings.json(confirmed/rejected fields) · the catalog must load 14 skills, 0 invalid, - launches ARIA from an unrelated working directory to guarantee the data paths are anchored to the project root,
- fails if any
__pycache__/*.pycsneaks back into the index.
- 1.0.0 · rule engine, fuzzy matching, persistent memory, feedback learning, facts, reflection, web search, GUI.
- 1.1.0 · reliability pass: fixed the reflection engine comparing an
exchange with itself, the web-search routing that never fired, an
eval()DoS in the calculator, the hard tkinter dependency, and CWD-relative data paths. - 1.2.0 · move to a declarative architecture: skill catalog
(
skills/*/SKILL.md), intent router with proof, memory discipline (no secrets ever, bounds, honored erasure), output masking and a security self-audit (--audit),--statscommand, hardened console and GUI. - 1.3.0 · catalog grew to 14 skills: task list, change journal, charter inspector, text statistics, CSV preview, unit conversions, world clock, calendar, local random values; reflection on its own line for multi-line answers.
- 1.3.1 · interface pass: branded console banner + integrated
aide/stats/auditcommands (routing traces only in--debug), GUI restyled to the brand (colors, routing status bar, chat clearing). See CHANGELOG.md.
- Richer fact schemas (age, city, preferences) + follow-up reasoning
- Multi-shot learning: keep several candidate answers per rule
- Accent/diacritics normalization before fuzzy matching
- Export/import of a memory as shareable "knowledge pack"
-
critical-risk skills: explicit confirmation before execution - Voice input via
speech_recognition(optional extra)
Issues and PRs are welcome · see CONTRIBUTING.md. Templates are provided for bugs, ideas and pull requests.
MIT · © 2026 Mr Jackson.

