Without an embedding API

Maybe you can’t reach OpenAI from your network. Maybe you don’t want to pay for embeddings on every write. Maybe you’re air-gapped. Maybe you just want to try Mnemosyne with the absolute minimum stack.

Mode A works. Mnemosyne ships with no embedding default and recall gracefully runs FTS-only when no embedder is configured.

How recall works in Mode A

Query
  ↓
Tier 3 cold path
  ↓
Vector stages → SKIPPED (no embedder)
FTS over text_lemmatized → ALWAYS RUNS
Trust ladder + recency + worth → ALWAYS RUNS
Compose top-K → return

The pipeline still ranks results by recency × worth × pin-status × FTS score. You lose semantic similarity but you keep textual similarity and all of the cognitive layer.

What you can still do in Mode A

• Create / read / update / pin / forget facts.
• Bitemporal time-travel queries (asOf).
• Entity graph and 1-hop expansion.
• All decisions, episodes, citations, audit.
• All MCP tools.
• All zero-LLM features (consensus, composition, meta-cognition, anticipation). They are SQL-based; embeddings aren’t required.

What you give up

• Cross-phrasing similarity. A query for “morning beverage” won’t match a fact about “coffee” unless the words overlap.
• HNSW speed. FTS over tsvector is slower than HNSW at large scale.

When you should pick Mode A

• Privacy-sensitive deployments: nothing leaves the machine.
• Cost-sensitive deployments: zero per-write cost.
• CI / testing: deterministic, fast, no external calls.
• Air-gapped intranet: no outbound network needed.

Switching to Mode B later

When you want vector recall, configure an embedder by setting the provider env vars and restarting:

MNEMO_LLM_PROVIDER=openai
MNEMO_LLM_API_KEY=sk-...

Future writes embed inline as they’re created, and reads start using the HNSW index as soon as rows have vectors.