Aura from Auralith: Compile documents into knowledge bases with 3-tier agent memory

[rtalabsai]

Built an extension that gives Gemini CLI persistent memory and instant document search.

What it does

• Compile any folder of documents (PDFs, DOCX, code, spreadsheets, markdown — 60+ formats) into a single queryable .aura knowledge base
• Query that knowledge base with natural language — instant RAG retrieval
• Remember context across sessions with a 3-tier Memory OS (pad, episodic, fact)

All processing runs locally on your machine. No data leaves your device. No API keys needed.

Install

pip install auralith-aura
gemini extensions install [https://github.com/Auralith-Inc/aura-gemini-cli](https://github.com/Auralith-Inc/aura-gemini-cli)

[aniruddhaadak80]

From my point of view, the local first angle here is compelling, especially for people who want better memory and document retrieval without sending more context to remote services. The part I would want made especially clear is provenance and trust boundaries inside the memory model, because once persistent memory exists users will want to know what came from their documents, what was inferred, and how to remove or reset it cleanly. That transparency matters almost as much as the retrieval quality.

[rtalabsai]

Great point, provenance and trust boundaries are exactly why Aura uses a tiered architecture instead of a single memory bucket.

Every memory entry stores its source (agent, user, or system), namespace (pad/episodic/fact), timestamp, and session_id — so you always know what came from where and when.

For reset/removal:

memory.prune_shards(before_date="2026-01-01"): prune by date
memory.prune_shards(shard_ids=["specific_shard"]): delete specific shards
memory.show_usage(): see exactly what's stored per tier
Or just delete ~/.aura/memory/ to wipe everything clean
Nothing is inferred or synthesized, memory only contains exactly what was written via write().

If the agent writes "user prefers dark mode," that exact string is what's stored and retrieved. No hidden embeddings, no derived data.

Good call on making this clearer in the docs, we'll add a provenance section.

Thanks for the feedback.