Akno

Akno is a memory layer for agents built on a Markdown knowledge base you own.

An agent can search, answer from, and deliberately update the same files you edit in Obsidian, vim, or any other editor. Markdown remains the source of truth. Akno adds rebuildable search projections plus durable undo, plan, receipt, and recovery state in SQLite.

Akno is for personal agents that need continuity across conversations without turning memory into an opaque chat-provider feature. Its memory stays inspectable, citable, reversible, and portable.

Status: Akno is in active pre-1.0 development and is used on a real personal knowledge base. The core reading, writing, ingestion, and maintenance workflows described below are implemented today; interfaces may still change between minor releases.

What works today

Area	Current behavior	Important boundary
Retrieval	Hybrid recall, full-page read, graph traversal, bounded agent context, and verified answer	Results cite page lines or document pages; empty, degraded, and unavailable remain distinct
Memory capture	Exact write, conversational remember, and replay-safe keyed retain with automatic extraction and routing	Writes are journalled and undoable; retained support can be corrected or retracted by source revision
Semantic memory	Current facts, history, plans, attributed reports, questions, and discussion are distinguished during retrieval	A proposal or reported claim is useful context without silently becoming a current fact
Time	timeline combines authored events, retained facts, plans, deadlines, and document dates against an explicit clock	Past, present, and future remain queryable without flattening every date into the same kind of fact
Documents	Files can be indexed as evidence, searched immediately, and later adopted into canonical pages	Searchable source material is not automatically promoted to canonical knowledge
Derived memory	Evidence-backed observations are co-located on admitted subject pages; reflected principles form a separate level	Observations keep exact leaf-fact lineage and become ineligible when their support changes
Maintenance	The seven-phase dream cycle detects conflicts, proposes exact repairs, organizes managed facts, and performs bounded synthesis	Scheduled maintenance defaults to audit; review and autonomous modes still use sealed plans and independent verification
Runtime	A single-writer service watches the knowledge base on macOS or Linux, with nightly launchd/systemd scheduling	Akno assumes one owner or one trusted shared owner; it is not a multi-tenant permission system

Why Akno

• Files stay authoritative. Indexing leaves the set and bytes of source files unchanged by default.
• Answers stay grounded. Retrieval returns evidence locators, and generated answers are verified separately against the evidence they cite.
• Absence stays honest. A clean miss is different from a failed model, unavailable document extractor, or degraded ranking path.
• Evidence is not automatically belief. Pages, source documents, observations, and ignored material have different roles and retrieval policies.
• Writes are bounded. Mutations are journalled; maintenance plans exact diffs before a human or separate curator decides what may apply.
• Organization never blocks retrieval. An orphan document is searchable before Akno finds or creates its canonical home.

Akno does not replace your editor, backup system, or judgment. It only knows what the indexed files and documents contain.

Quick start

Akno supports macOS and Linux and requires Node 22.18 or newer.

npm install -g @tenphi/akno
akno init

Guided setup asks for the notes folder, model strategy, folder roles, and maintenance authority. It can configure the benchmark-qualified OpenAI preset, a model-free lexical setup, or preserve an existing specialist setup. Every optional follow-up—indexing, a first recall, and background-service installation—requires its own confirmation.

For the OpenAI preset, provide the credential through the environment. Akno stores only the variable name:

export AKNO_OPENAI_API_KEY="..."
akno init

Then index and inspect the knowledge base:

akno index
akno doctor
akno recall "How long is the Zephyr QX-100 warranty?"
akno recall "What is planned for the Zephyr QX-100?" --memory-view planning

Markdown indexing needs no extra system tools. Document extraction uses native macOS frameworks; on Linux, PDF, OCR, and office-file support use Poppler, Tesseract, and LibreOffice respectively. Missing tools degrade only the affected extraction path and are reported by akno doctor.

Akno keeps requests on the selected model provider's configured origin: redirects cannot cross origins, and a failed provider is never silently replaced by another configured provider.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗