v2.12.0 — Phase 3 Facts + Confidence, embedding fallback, Dreamer rehab

Three meaningful additions bundled in one release. Server app version bumps from 0.7.0 to 0.8.0; package version 2.11.3 → 2.12.0.

Phase 3 — Structured Facts Table + Three-State Confidence

The gap this closes: Mnemo today treats every memory as text+tags retrieved by FTS5 or vector similarity. That's the right tool for fuzzy recall over prose. It's the wrong tool for key-value lookups ("what is X's location?") that should update in place rather than accumulate as parallel memory chunks.

What ships:

• New SQLite store at ~/.agentb/facts.sqlite (shared global, WAL mode, schema auto-creates on connect)
• facts table — composite PK (entity, attribute), one current value per pair
• fact_history table — append-only audit log of every change
• Three-state confidence (verified > high_probability > false) with promotion ladder enforcement
• Six new HTTP routes (/facts/{e}/{a}, /facts?filters, POST /facts, POST /facts/demote, /facts/history, /facts/contradictions)
• Four new MCP bridge tools (mnemo_fact_get / mnemo_fact_query / mnemo_fact_save / mnemo_fact_demote)
• Evidence source prefix convention (memory:<id>, commit:<sha>, statement:<who>, etc.)
• 23 unit tests covering the full promotion-ladder semantics

Dreamer Stage 0.5 — automated fact extraction: nightly Dreamer cron calls the LLM with a strict conservative-only-direct-statements prompt to extract (entity, attribute, value) triples from new memories and POST them to /facts with confidence='high_probability'. Auto-capture entries are filtered before extraction (they're tool-call logs, not stated facts).

Contradiction notification: when extraction conflicts with an existing verified fact, the spec's promotion ladder correctly rejects the overwrite — but silent rejection means contradictions pile up. v2.12.0 batches all per-run conflicts and posts at end of Dreamer run to two optional channels:

• Bus message via HTTP (requires MNEMO_DREAM_BUS_URL + MNEMO_DREAM_BUS_FROM + MNEMO_DREAM_BUS_TARGETS env vars)
• Discord webhook (MNEMO_DREAM_DISCORD_WEBHOOK env var)

Both opt-in, both gracefully skip when unset. One batched message per cron run, never per-contradiction.

Stage 0.5 adds ~1 LLM call per agent per night, doubling per-night dream cost from $0.0013 to $0.003.

Embedding hosted fallback (Google Matryoshka)

GoogleEmbedding.embed() now honors extra.output_dimensionality, enabling a fully working hosted fallback for the local Ollama embedding primary. Google's gemini-embedding-001 outputs 3072 dims natively; Matryoshka truncation coerces to the locked sqlite-vec store width (768) without dim-guard trips. See agentb.yaml.example for the config shape. Closes the "if Ollama dies, every Mnemo recall 500s" gap.

Dreamer pipeline rehab

mnemo-dream.py had been silently dark for a week due to three cascading failures (token explosion, env file gap during v2.10.0 cutover, path mismatch). v2.12.0 fixes all three:

• Path migration: harvest walks ~/.agentb/agents/<agent>/memory/ (current layout). Agent auto-discovery enumerates the directory at runtime; MNEMO_DREAM_AGENTS env var pins a specific list when needed.
• Two-stage map-reduce synthesis: per-agent partial summaries first, then joint cross-agent rollup. Per-call token usage bounded; 1M context limit never approached.

Upgrade notes

• pip install --upgrade mnemo-cortex and restart the server. Schema auto-creates; no migration step.
• New env vars (all optional):
  • MNEMO_DREAM_BUS_URL / MNEMO_DREAM_BUS_FROM / MNEMO_DREAM_BUS_TARGETS — contradiction-to-bus notification
  • MNEMO_DREAM_DISCORD_WEBHOOK — contradiction-to-Discord notification
  • DREAM_SKIP_FACTS — set to 1 to disable Stage 0.5 (debug)
• See agentb.yaml.example for the full v0.8.0 config shape including the new embedding fallback pattern.
• See CHANGELOG.md v2.12.0 section for full detail.