Aura Research

Turn raw research into a living wiki your LLM agents can read, query, and enhance.

---

Aura Research is an LLM-powered research knowledge base that compiles raw documents into a structured markdown wiki. Drop your papers, articles, data, and notes into a folder — the LLM reads everything, builds a navigable wiki with summaries and concept articles, and then answers your research questions using that compiled knowledge.

Built on Aura Core for document compilation (60+ formats) and the three-tier Memory OS for persistent agent memory across sessions.

Quick Start

# Install
pip install 'aura-research[openai]'

# Set your API key
export OPENAI_API_KEY=sk-...

# Initialize a project
research init my-project
cd my-project

# Drop your documents in raw/
cp ~/papers/*.pdf raw/

# Ingest and compile
research ingest raw/
research compile

# Ask questions
research query "what are the key findings across all papers?"

# Search the wiki
research search "attention mechanism"

# Check wiki health
research lint

Agent-Native Mode (no API key)

If you're already using an AI coding agent (Claude Code, Codex, Gemini CLI, Cursor, etc.), you don't need an API key. The agent IS the LLM:

# In your AI agent's terminal:
research init my-project
# Copy documents to raw/
research ingest raw/

# The agent reads the docs and writes wiki articles directly
# (it's an LLM — it doesn't need to call another one)

research build          # compile wiki/ → wiki.aura
research search "topic" # search the wiki
research memory show    # see what the agent remembers

The API mode (research compile, research query) exists for headless/batch use when no agent is at the keyboard.

How It Works

Raw Documents  ──→  Aura Core (.aura)  ──→  LLM Compiler  ──→  Markdown Wiki
  papers/              compiled &              generates          wiki/
  articles/            indexed                 summaries,         ├── _index.md
  data/                (60+ formats)           concepts,          ├── concepts/
  code/                                        backlinks          ├── sources/
                                                                  └── queries/
                              ↕
                        Memory OS
                   /pad  /episodic  /fact
                   (persistent agent memory)

1. Ingest — Compile your raw documents into a searchable .aura archive using Aura Core
2. Compile — LLM reads all sources and generates a structured wiki: per-source summaries, cross-cutting concept articles, master index, and executive summary
3. Query — Ask questions against the wiki. The LLM uses wiki context + Memory OS facts + optional web search to give you thorough, cited answers
4. Remember — Memory OS automatically stores key findings (/fact), session logs (/episodic), and working notes (/pad) — so the agent never starts cold

Commands

Command	Description
research init [dir]	Initialize a new project
research ingest <dir>	Ingest raw documents into .aura archive
research ingest <dir> --watch	Watch directory and auto-re-ingest on changes
research compile	Compile wiki using LLM API (needs API key)
research compile --full	Full recompile (ignore cache)
research build	Build wiki.aura from wiki/ markdown (no LLM needed)
research query "..."	Ask a research question (needs API key)
research query "..." --save	Ask and save the answer to wiki/queries/
research query "..." --no-web	Ask without web search
research search "..."	Keyword search across wiki articles
research lint	Run wiki health checks
research lint --ai	Health checks + AI-powered analysis
research status	Show knowledge base statistics
research memory show	Full overview of all 3 memory tiers
research memory show --tier fact	Overview filtered to one tier
research memory usage	Show Memory OS storage
research memory query "..."	Search agent memory
research memory write <tier> "..."	Manually write to memory (pad/episodic/fact)
research memory list	List memory shards
research memory prune --before DATE	Prune old memory entries

LLM Providers

Supports OpenAI, Anthropic, and Google Gemini. Install the one you prefer:

pip install 'aura-research[openai]'       # OpenAI models (default)
pip install 'aura-research[anthropic]'    # Anthropic models
pip install 'aura-research[gemini]'       # Google Gemini models
pip install 'aura-research[all]'          # Everything

Configure via environment variables or research.yaml:

llm:
  provider: openai          # openai, anthropic, or gemini
  model: gpt-5.4-instant    # override default model
  temperature: 0.3

memory:
  enabled: true
  auto_write: true           # agent writes to memory automatically

web_search:
  enabled: true
  max_results: 5

watch:
  enabled: false
  interval: 5               # seconds between checks

Memory OS

Aura's three-tier Memory OS v2.1 gives the agent persistent memory across sessions — so it never starts cold:

Tier	What's Stored	Persistence
/pad	Working notes, draft observations	Transient — scratch space
/episodic	Session logs, what was compiled/queried	Auto-archived
/fact	Key findings, verified observations	Persistent — survives indefinitely

How It Operates

Memory OS works both autonomously and manually:

• Autonomous: During research compile, the LLM auto-extracts key facts and writes them to /fact. After every compile/query session, an episodic log → /episodic. Controlled by auto_write: true in config.
• Manual: Write directly with research memory write <tier> "content" whenever you (or the agent) want to persist something.

v2.1 Features

Feature	What It Does
Entry deduplication	Prevents writing the same fact twice (SimHash fuzzy matching)
Temporal decay	Recent memories score higher in queries — older context naturally fades
Bloom filters	Skip irrelevant shards during search — fast even with thousands of entries
Append-only	Old entries are never overwritten — new ones are added alongside them
Tiered priority	Facts > episodic > pad when returning query results

Examples

# Write a verified fact
research memory write fact "The model achieves 94.2% accuracy on the test set"

# Log what you did this session
research memory write episodic "Analyzed training curves, found overfitting at epoch 12"

# Jot a working note
research memory write pad "TODO: re-run experiment with lower learning rate"

# Search memory by keyword
research memory query "accuracy"

# Full overview — see everything across all 3 tiers
research memory show

# Filter to just facts
research memory show --tier fact

# Storage usage
research memory usage

Web Search

During research query, the agent can search the web to supplement wiki answers with current information. This is enabled by default and uses DuckDuckGo (no API key required).

pip install 'aura-research[search]'

# Query with web search (default)
research query "latest advances in attention mechanisms"

# Query without web search
research query "what does our data show" --no-web

Watch Mode

Auto-detect new files and re-ingest:

pip install 'aura-research[watch]'

# Watch for changes (uses watchdog if installed, falls back to polling)
research ingest ./papers --watch

Wiki Output

The compiled wiki lives in two places:

wiki.aura — The primary artifact. An .aura archive containing all wiki articles, optimized for agent RAG retrieval. Token-efficient — agents read only what's relevant.

wiki/ — Markdown export for human browsing. Open in Obsidian, VS Code, GitHub, or any markdown viewer:

.research/
├── knowledge.aura      ← Raw ingested documents
└── wiki.aura           ← Compiled wiki (agent reads from here)

wiki/
├── _index.md           ← Master index with links to all articles
├── _summary.md         ← Executive summary of the knowledge base
├── concepts/           ← Cross-cutting concept articles
│   ├── attention.md
│   ├── tokenization.md
│   └── ...
├── sources/            ← Per-source summary articles
│   ├── vaswani2017.md
│   ├── devlin2019.md
│   └── ...
└── queries/            ← Saved Q&A responses
    └── ...

After editing wiki articles (or having the agent write them), run research build to recompile wiki.aura.

License

Apache License 2.0 — see LICENSE.

Links

• Aura Core: github.com/Rtalabs-ai/aura-core — the universal context compiler
• Rta Labs: rtalabs.org