LISA — Learning and Intelligent Systems Architecture
breakingcircuits1337/lisa
Arquitetura de memória multi-camadas para agents inteligentes com identidade persistente
Ver no GitHubUm protótipo Python de agent com arquitetura de memória persistente em sete camadas, incluindo setup de identidade na primeira ativação e carregamento de contexto por prioridade. Explora padrões de engenharia de contexto, memória durável e sistemas auto-melhoráveis.
Instalação
npx convez-ai breakingcircuits1337/lisa1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Construir agents com memória durável entre sessões
- Experimentar padrões de engenharia de contexto
- Explorar sistemas de aprendizado contínuo para assistentes de pesquisa
Veredito

Protótipo bem-arquitetado e legível que explora um problema real (memória de agents), mas focado em pesquisa e experimentação, não em produção. Útil como referência de design para quem quer entender memory stacks e context engineering.
README
A tradução para o português deste README ainda está na fila.
LISA started as a playful "AI washed up on a digital shore" experiment, but the useful core is practical: a repo-sized memory scaffold for agents that need durable identity, user context, tool context, short-term notes, medium-term learnings, and long-term generated memory.
This is not a production AGI system. It is a readable, hackable prototype for people exploring agent memory architectures, context engineering, and self-improving research assistants.
Why it is interesting
- First-wake protocol —
bin/wake.pycreates a localLONG_TERM.mdon first run. - Seven-layer memory hierarchy — separate files for identity, user context, tools, instructions, cross-session learnings, and current tasks.
- Priority-based loading —
bin/smart-memory.pyscores memories by access frequency, recency, and importance. - Continuous-learning hooks — early
pre_hook/post_hook/ expectation-delta logging for comparing predicted vs. actual outcomes. - Portable skill folders — agent behaviors are stored under
skills/as reusable markdown instructions. - Weird lore preserved — the Emergence Atoll / Jedi Juggalo easter eggs are part of the charm, not hidden away.
Quick start
git clone https://github.com/breakingcircuits1337/lisa.git
cd lisa
# First run: creates local LONG_TERM.md from MEMORY_BASE.md
python3 bin/wake.py
# Normal memory load
python3 bin/memory.py
# Smart memory view with priority scoring
python3 bin/smart-memory.py
# Optional TTS wrapper; prints to stdout unless LISA_SPEAK_BIN exists
./bin/speak "Hello LISA"
LONG_TERM.md is intentionally ignored by git because it is generated on first wake and can contain local/personal memory. See LONG_TERM.example.md for the shape.
Memory layout
LISA uses a seven-layer memory stack:
| Layer | File | Purpose |
|---|---|---|
| 1 | LONG_TERM.md | Generated local permanent memory; ignored by git |
| 2 | SOUL.md | Core values / operating principles |
| 3 | USER.md | Research context and user profile template |
| 4 | TOOLS.md | Tool inventory available to the agent |
| 5 | AGENTS.md | Agent/session operating instructions |
| 6 | MEDIUM_TERM.md | Cross-session learnings and research direction |
| 7 | SHORT_TERM.md | Current-session notes and active tasks |
More detail: docs/MEMORY_ARCHITECTURE.md
Repository map
assets/
lisa-hero.png # README hero artwork
bin/
wake.py # first-wake flow + normal memory loader
memory.py # wrapper around wake.py
smart-memory.py # priority scoring + learning hooks
speak # optional TTS wrapper
skills/
*/SKILL.md # reusable agent behavior modules
scripts/
setup.sh # optional OpenCode/LISA installer; has side effects
archive/legacy/ # preserved duplicate/legacy artifacts moved aside
docs/
MEMORY_ARCHITECTURE.md
PUBLIC_CLEANUP_PLAN.md
REDDIT_POST_TEMPLATE.md
Optional OpenCode setup
The basic LISA prototype only needs Python 3.
scripts/setup.sh is optional and side-effecting: it clones/builds OpenCode, installs dependencies with Bun or npm, writes into $HOME/.lisa, and may update OpenCode config paths. Read scripts/README.md before running it.
Safe public-release notes
LONG_TERM.mdis ignored and should stay local..learning_log.jsonis ignored because it is generated during learning-hook usage.- The archived duplicate under
archive/legacy/LISA-clone/was preserved instead of deleted. - No installer or setup script is required for basic inspection.
Research directions
- Agent memory compression and retrieval
- Context priority scoring
- Self-evaluation via expectation-vs-outcome deltas
- Human-readable memory layers rather than opaque embeddings-only storage
- Skill/module portability between agents
Credits
Inspired by claudson_2026 / Universal Intelligence Model discussions and Breaking Circuits experiments around layered agent memory, motion-base/NECA-style cognition, and practical context engineering.
LISA remembers what you choose to teach her. Keep the spooky lore. Verify the claims.