Default keyword recall is unchanged. Configure Supabase to upgrade to cosine similarity on real embeddings.
ar setup supabase # interactive wizard
ar setup supabase --migrate # apply pgvector migration
ar setup supabase --backfill # re-embed all local memoriesSupports OpenAI text-embedding-3-small (1536 dims) and Voyage voyage-3-lite (512 dims, zero-padded to 1536). Local files remain the source of truth — Supabase is a derived read index.
Graceful degradation: if Supabase is unreachable, recall() falls back to local keyword search silently.
Four new tools added to the MCP server:
| Tool | What it does |
|---|---|
project_board |
Status board across all projects (equivalent to /arstatus) |
project_status |
Deep status for a single project |
bootstrap_scan |
Scan machine for existing git repos, Claude AutoMemory, CLAUDE.md files |
bootstrap_import |
Import discovered projects into AgentRecall |
A new decisions room is now included in the default palace: DEFAULT_PALACE_ROOMS. Tracks decision trails with prior/posterior tracking, evidence chains, and outcomes. Bayesian-inspired audit trail for major decisions.
Projects created before v3.4.0 will have the decisions room auto-created on next palace initialization.
Awareness insight cap increased from 15 to 20. Existing awareness files are unaffected.
No breaking changes. All existing data, palace rooms, journals, and awareness files are compatible with v3.4.0.
If you use the Supabase backend, run ar setup supabase --backfill to embed existing memories.
All packages updated to 3.4.0:
agent-recall-mcp@3.4.0agent-recall-sdk@3.4.0agent-recall-cli@3.4.0agent-recall-core@3.4.0