Every time Claude Code, Cursor, or Codex starts a fresh session, it starts from zero — your preferences, your past decisions, and your codebase's quirks are gone. Rivera is a memory CLI that persists all of that across sessions and across tools, backed by Rivera, a semantic memory engine with exact (non-approximate) vector search and zero indexing delay: store a memory and it is searchable the same millisecond.

$ rivera remember "cua-driver Rust build needs DEVELOPER_DIR pointing at full Xcode" --type fact
Memory stored successfully!  Type: fact | Confidence: 0.95

$ rivera recall "how do I build the rust driver"
→ cua-driver Rust build needs DEVELOPER_DIR pointing at full Xcode   (score 0.39)

$ rivera answer "what does the rust build need?"
→ The build requires full Xcode via DEVELOPER_DIR (Source: chunk 1).

Three primitives

Memories are typed (13 categories: fact, preference, decision, goal, instruction, learning, error, …) and carry confidence and provenance metadata, so an explicit user statement never gets confused with an inferred hunch.

Quickstart

# 1. Install
pip install rivera-cli

# 2. Get a free API key at https://api.wirtel.ca/signup
#    (free plan: 2,000 requests + 200 GenAI answers / month)
export RIVERA_API_KEY="rv_..."

# 3. Create your agent and go
rivera agent create my-agent
rivera remember "User prefers concise answers" --type preference
rivera recall "communication style"
rivera answer "what did we decide about the database schema?"

Configuration lives in ~/.rivera/ (.env for credentials, config.yaml for settings). Point at a different Rivera deployment with RIVERA_BASE_URL.

Agent integrations

Connect Rivera to your coding agent so memory works automatically — context injected at session start, durable decisions captured as you work:

rivera connect claude-code    # also: cursor, codex, windsurf, cline, continue, ...

More than a CLI

• Local REST API + Web UI — rivera serve / rivera ui
• Batch ingestion — rivera remember --batch memories.json, --from-conversation to extract facts from chat logs, rivera upload for PDF/DOCX/CSV/MD files
• Memory hygiene — rivera conflicts detects contradictions; rivera daily-summary digests what changed; rivera edit / rivera forget for corrections
• Project sync — rivera memory sync writes a MEMORY.md snapshot into your repo

How it works

rivera CLI ──HTTPS──▶ Rivera (api.wirtel.ca)
                      ├─ exact cosine search over pgvector (no ANN, deterministic)
                      ├─ OpenAI embeddings (text-embedding-3-small)
                      └─ grounded answers (gpt-4o-mini) with citations

Your memories live in your Rivera account — per-tenant isolation, API keys hashed at rest, revocable from the console. Self-hosting Rivera is possible too: the backend is a standard FastAPI + Postgres/pgvector service.

Acknowledgments

Rivera began as a fork of memanto (MIT, EdgeAI Innovations Inc.) and preserves its excellent typed-memory model and CLI ergonomics. The backend, retrieval engine, auth, and cloud service are Rivera — an independent implementation. See LICENSE.

License

MIT