Memory over find.

One contract for every memory scope

The swarm already runs real memory infrastructure — remote Qdrant Cloud vector memory, a code/docs impact graph, and the hive's Obsidian knowledge vault. Mnemosyne doesn't reinvent any of it — it unifies what already runs behind a single, escalating, provenance-tracked API.

Meta, broad, to file, raw

An ordered, escalating stack. Backends are pluggable per slot — Qdrant is the default vector backend, Obsidian the default meta store, but neither is a hard dependency.

Component & flow

Full diagrams (including the recall() sequence flow) live in docs/architecture.md.

flowchart TB
  subgraph pantheon["Pantheon gods (callers)"]
    minerva["Minerva
planner"]
    argus["Argus
metrics"]
    swarm["swarm agents"]
  end

  subgraph mnemosyne["Mnemosyne — memory god"]
    api["recall(query, scope, intent)
remember(content, scope, layer?)"]
    router["layer router + escalation
(narrow to broad, merge + rank)"]
    prov["provenance stamping
(7 fields per hit)"]
    idx["continuous indexing
(Multica-native schedule)"]
    api --> router --> prov
  end

  subgraph layers["Layer stack (pluggable slots)"]
    meta["meta - Obsidian vault"]
    ent["enterprise"]
    proj["project"]
    cg["code-graph"]
    vec["vector - Qdrant Cloud"]
    file["file - grep (loud floor)"]
  end

  minerva --> api
  argus --> api
  swarm --> api

  router --> meta
  router --> ent
  router --> proj
  router --> cg
  router --> vec
  router --> file

  idx -.keeps fresh.-> vec
  idx -.keeps fresh.-> cg
  idx -.keeps fresh.-> meta

  vec -->|wraps| sm[("swarm-memory
+ Qdrant Cloud")]
  cg -->|wraps| sm

  api -.decision + metric record.-> argus

  consus["Consus / Janus
(read model)"] --> api
      

Clone, install, run the tests

gh repo clone mdostal/mnemosyne
cd mnemosyne
npm install

npm test              # full suite
npm run test:e2e      # Minerva-style end-to-end integration test
npm start              # run the production recall/remember service (:8477)

See README.md for the full command reference, docs/http-api.md for the MnemosyneClient HTTP API, and hooks/README.md to wire the pre-recall/post-remember hooks into an agent runner.

One god, one capability slot

• Hostpantheon-v2 — the core host that assembles gods behind shared contracts.
• SubstratePlanned and executed on Multica with the plugin-hive SDLC.
• CallersMinerva (planner) and swarm agents are the primary recall callers.
• Read modelConsus / Janus provide the human view — browse layers, trace provenance, spot stale scopes.
• Backend configVesta owns which backend fills each layer slot.
• MetricsArgus / Metis receive the decision + metric record from every recall/write.
• Builds onswarm-memory — Qdrant-backed semantic memory + code/docs impact graph, wrapped, not rewritten.