The concept map

Every explanation of The Athanor traverses one graph. Only the depth changes; the definitions do not.

The Athanor is infrastructure that gives AI tools bounded, attributed access to project knowledge and, when needed, durable governed continuity that can survive closed sessions and changing model processes.

One system, any depth

Read the diagram source

flowchart TB
    ATH[The Athanor] --> CORE[Core contracts]
    ATH --> HOUSE[House]
    HOUSE --> ROOM[Rooms]
    CORE --> RET[Retrieval]
    CORE --> CONT[Continuity]
    CORE --> ROUTE[Worker routing]
    RET --> VAULT[Vault]
    RET --> AKASHA[AKASHA]
    VAULT --> FILES[Attributed project files]
    AKASHA --> AUTH[Typed PostgreSQL authority]
    AUTH --> CANON[Canon]
    AUTH --> MEMORY[Memory and lessons]
    AKASHA --> GIGA[GIGA]
    GIGA --> HIP[Hippocampus Stage 1]
    GIGA --> STR[Striatum process-lesson slice]
    HIP --> CAND[Non-authoritative candidates]
    CONT --> BOAT[Paper Boats]
    ROUTE --> LANES[Worker lanes and familiars]

Words that keep their meaning

The Athanor

The platform and core architecture. It defines continuity, retrieval, authority, routing, and extension contracts. Its current supported public harness is OMP. Provider-neutral core contracts do not imply that other harness adapters are already supported.

House

One sovereign continuity domain. A House can contain multiple rooms and shared work knowledge without erasing room boundaries. A House is not required merely to search a project corpus through Vault.

Room

A bounded line of identity, relationship, context, and continuity within a House. Rooms prevent every memory and interaction from becoming one undifferentiated assistant history.

Vault

The lightweight file-backed retrieval profile. It searches configured Markdown, JSON, JSONL, and plain-text roots with exact-content and field-aware BM25F lanes. It returns bounded attributed excerpts and requires no database, embeddings, or GPU.

AKASHA

The PostgreSQL-backed authority and retrieval profile. It adds pgvector, pg_trgm, compatible local embeddings, typed stores, chronology, supersession, taxonomy, semantic and structured retrieval, and the substrate used by GIGA.

Canon

Current authoritative truth in AKASHA. Canon outranks loose memory. A retrieval score cannot promote a record into canon.

Memory and lessons

Memory carries durable continuity with provenance and lifecycle. Lessons are typed reusable knowledge: coding, project, writing, design, and audio. They have their own eligibility and guarded-write contracts.

GIGA

Grounded Indexing and Generative Annotation: the cognitive layer above AKASHA, not another storage profile. Hippocampus Stage 1 produces grounded, non-authoritative candidates for review. Striatum's current slice braids process-shape coding lessons from deterministic process triggers; the state-conditioned selector remains planned. Candidates do not become truth by existing.

Organs

Named deterministic tools of a House. The supported OMP adapter currently mounts 26 across memory, retrieval, lessons, continuity, counsel, routing, design-system catalogues, GIGA review, and House configuration. An organ is not an autonomous agent.

Paper Boat

A compact living handoff from one closed session to the next waking session in the same room. It is continuity, not the entire memory map.

The short path for project work

Project work begins with the problem the operator can see:

Your AI tool only works with the context it has. The Athanor can search one or several project corpora, return the relevant passages with their sources, and preserve important decisions when you opt into its durable memory profile.

The path stays short:

Read the diagram source

flowchart LR
    PROJECTS[Your projects] --> V[Vault]
    QUESTION[Your task] --> V
    V --> SOURCES[Relevant excerpts and sources]
    SOURCES --> TOOL[Your AI work tool]
    TOOL --> RESULT[Work you can inspect]

AKASHA enters only when the work needs durable typed decisions, semantic recall, lessons, chronology, or a larger governed archive. None of those concepts are required to understand the immediate value of attributed project context.

The architectural path

The architectural distinction is:

The Athanor separates model judgment from continuity, retrieval, authority, provenance, and repeated deterministic cognition. The active model is a replaceable body operating inside those contracts, not the database of truth.

That path has five parts:

1. Vault supplies attributed project retrieval without infrastructure weight.
2. AKASHA supplies typed PostgreSQL authority and hybrid retrieval.
3. Canon, memory, lessons, candidates, and counsel have different authority.
4. GIGA may propose; authorized lifecycle operations decide what becomes durable.
5. Repeated known cognitive work should become an organ rather than another prompt ritual.

Continue with For latent-space explorers.

The operational traversal

The operational traversal is explicit:

Need current project facts?
  → recall through Vault or AKASHA
  → preserve source, heading/record, authority, and selection reasons

Need durable continuity?
  → remember through the authorized store
  → make the body stand alone; source paths are provenance, not substance

Need reusable practice?
  → query the typed lesson store before risky work
  → respect project, type, stage, register, and scope eligibility

Need candidate review?
  → treat GIGA output as a proposal
  → review and promote explicitly; never cite a candidate as authority

Need work delegation?
  → resolve a bounded worker lane or familiar
  → inspect the packet; the harness performs the spawn

The agent should read Usage for operation names and Architecture for authority boundaries before teaching behavior it has not verified.

Thirty seconds

The Athanor makes AI tools more reliable on real projects by controlling how they receive context. Vault can search local project files and return bounded excerpts with exact attribution without needing a database or GPU. AKASHA adds PostgreSQL-backed memory, typed lessons, semantic retrieval, corrections, and governed cognitive workers. The model can change; the sources, authority, and continuity contracts remain inspectable.

Two minutes

Most AI tools either forget between sessions or recover context by dumping files, transcript summaries, and vector-search results into a prompt. That can retrieve useful text, but it does not distinguish current truth from stale memory or a machine-generated suggestion.

The Athanor supplies that missing structure. Its Vault profile is the small path: point it at one or several Markdown, JSON, JSONL, or text corpora and it returns relevant bounded excerpts with source and match attribution. Its AKASHA profile adds a PostgreSQL authority layer, typed memories and lessons, semantic retrieval, chronology, supersession, and GIGA workers whose output is explicitly non-authoritative until reviewed.

A House is the sovereign continuity boundary; rooms keep identities and relationships separate inside it. Deterministic organs handle known mechanics such as retrieval, validation, lifecycle, and routing so models can spend judgment on ambiguity and novel work. Today the only supported release target is Windows x64 with OMP. The Host supports the read-only web operator surface at gui-prototype/. bun gui-prototype/serve.ts serves it and proxies Host reads over loopback. The Godot client is parked. Cingulate, OMEGA, and ANON remain roadmap claims.

What The Athanor is not

The Athanor is not:

• merely a chatbot memory plugin;
• a vector database with mythology around it;
• a multi-agent orchestration framework whose main purpose is spawning agents;
• a claim that a model process possesses metaphysical personal continuity;
• an enterprise knowledge platform already supporting teams and tenancy;
• a privacy layer that prevents model providers from seeing prompts;
• a cross-platform, one-click, harness-agnostic release today;
• a finished native application; the Godot client is parked;
• a system where GIGA candidates, retrieved memories, or Anamnesis counsel are automatically authoritative.

The distinctions below carry the architecture:

Status language

Use status words exactly:

• Current: used by the reference House now.
• Specified: a written contract exists; implementation may not.
• Planned: accepted roadmap direction.
• Research: an investigated possibility without a release promise.

For current labels, consult Planned Features. For measured claims, consult Evidence. For the supported boundary, consult Limitations.

Recommended reading by question