Mnemosyne supports multiple isolated knowledge graphs under one supervision tree. Each repository has its own MemoryStore process and GraphBackend state.
When to Use Multiple Repos
Use separate repos for project, tenant, user, environment, or domain boundaries that must not share knowledge.
{:ok, _} =
Mnemosyne.open_repo("project-alpha",
backend:
{Mnemosyne.GraphBackends.InMemory,
persistence:
{Mnemosyne.GraphBackends.Persistence.DETS,
path: "priv/memory/alpha.dets"}},
telemetry_labels: %{environment: :production, team: "search"}
)
{:ok, _} =
Mnemosyne.open_repo("project-beta",
backend: {Mnemosyne.GraphBackends.InMemory, []},
telemetry_labels: %{environment: :staging, team: "platform"}
)For repository cost observability, repo_id is the tenant identity. :telemetry_labels is an optional flat map with string or atom keys and scalar string, atom, number, or boolean values. Labels live only in the open MemoryStore; they are not stored in the graph backend. Supply them again when reopening a repo.
Do not put secrets in labels. Keep values bounded and low-cardinality: request IDs, user IDs, prompts, and unbounded source IDs create expensive or unusable metric dimensions.
Mnemosyne.list_repos/1 lists open IDs, and Mnemosyne.close_repo/2 closes one repository. Opening an existing ID returns RepoError with :already_open; using a missing ID returns NotFoundError for :repo.
Repo-Scoped Source IDs
A trajectory's source_id is stable within its target repo, not globally unique. These two complete trajectories have independent identity records:
trajectory = %Mnemosyne.Trajectory{
source_id: "task-42",
goal: "Explore caching",
steps: [
%{observation: "A cache miss occurred", action: "Inspected the key namespace"}
],
metadata: %{workflow: "diagnosis"}
}
{:ok, alpha_receipt} = Mnemosyne.ingest("project-alpha", trajectory)
{:ok, beta_receipt} = Mnemosyne.ingest("project-beta", trajectory)Within either repo, an equal retry returns that repo's exact original receipt and a different payload for "task-42" conflicts. Source records and graph content never cross repo boundaries.
All other operations also take a repo ID first:
{:ok, memories} = Mnemosyne.recall("project-alpha", "How does caching work?")
graph = Mnemosyne.get_graph("project-alpha")
:ok = Mnemosyne.delete_nodes("project-alpha", ["node-1"])
:ok = Mnemosyne.decay_nodes("project-alpha")Deleting or decaying nodes in a repo leaves its ingestion records intact.
Repository Cost Telemetry
Attach host telemetry handlers to the canonical terminal events. :stop records completed adapter calls, including returned adapter errors; :exception records raised, thrown, or exited calls. The :start event supports timing only and is not a usage or cost record.
:telemetry.attach_many(
"my-app.mnemosyne-costs",
[
[:mnemosyne, :model, :call, :stop],
[:mnemosyne, :model, :call, :exception]
],
&MyApp.Telemetry.handle_model_call/4,
nil
)Both terminal events include repo_id, nested labels, operation, optional source_id, kind, callback, adapter, step, requested_model, and status. Stop-event metadata also includes response_model and optional currency; exception-event metadata instead includes exception_kind, reason, and stacktrace. Terminal measurements include native duration, while stop measurements may also include recognized provider-reported token or cost fields. Flatten nested labels only in the host exporter when its metric backend requires it; Mnemosyne keeps them nested to avoid collisions with trusted fields.
Aggregate total_cost by repo_id in the host observability system. Do not infer a total from component costs, treat missing cost as zero, or double count existing LLM or embedding adapter events. Those adapter events remain backward-compatible but are non-canonical for per-repo cost aggregation.
Mnemosyne does not price, aggregate, persist costs, invoice, budget, or provide operation-level labels. It emits observability data, not a billing ledger. Multi-repo tenant grouping, billing ledgers, budgets, and operation-level custom labels are not provided.
Configuration Boundaries
Supervisor configuration supplies shared defaults. A repo can override config and adapters when opened. A complete ingestion can replace its config and adapters and pass pipeline-wide llm_opts; ordinary trajectory embeddings use config options, while per-call embedding_opts currently applies only to write-time intent merging. These execution choices do not affect payload identity. Recall uses the repository config and adapters.
Backend configuration is always per repo because each repo owns separate backend state.
Supervision Architecture
Mnemosyne.Supervisor (rest_for_one)
|-- RepoRegistry
|-- TaskSupervisor
|-- RepoSupervisor (DynamicSupervisor)
|-- MemoryStore "project-alpha"
|-- MemoryStore "project-beta"Ingestion, recall, write, and maintenance tasks use the shared TaskSupervisor; each repo's MemoryStore owns admission and backend state.
Multiple Supervisor Instances
Run independent Mnemosyne instances with custom names:
{Mnemosyne.Supervisor,
name: MyApp.WorkMemory,
config: work_config,
llm: work_llm,
embedding: work_embedding}
{Mnemosyne.Supervisor,
name: MyApp.PersonalMemory,
config: personal_config,
llm: personal_llm,
embedding: personal_embedding}Pass the matching supervisor to every operation:
Mnemosyne.open_repo("repo", backend: backend, supervisor: MyApp.WorkMemory)
Mnemosyne.ingest("repo", trajectory, supervisor: MyApp.WorkMemory)
Mnemosyne.recall("repo", "query", supervisor: MyApp.WorkMemory)