Question 1

How do I install Semvec?

Accepted Answer

pip install semvec for the core. pip install "semvec[cortex,coding,api]" for the full suite (multi-agent, coding-agent MCP server, FastAPI REST). Pre-built wheels ship for Linux (glibc + musl), macOS (Intel and Apple Silicon), and Windows x86-64. Python 3.10 through 3.14, stable ABI.

Question 2

How does Semvec compare to mem0?

Accepted Answer

Same job, different cost class. Semvec runs zero generative LLM calls at ingest; mem0 issues roughly 50 fact-extraction calls per stored turn. On a LOCOMO-scale workload that compounds to tens of millions of ingest tokens for mem0 versus zero for Semvec — an order of magnitude or more less total tokens. On LOCOMO conv-44 (675 turns, gpt-4o-mini, T=0), Semvec replays in ~3 min vs mem0's ~24.5 min (~8× faster). mem0 leads on the LOCOMO LLM-as-Judge headline by ~7 pp (0.669 vs 0.597); we publish both numbers. The cost-class gap lives in the ingest layer — invisible in per-turn comparisons because per-turn reader-context is roughly comparable, but dominant in any realistic total-token rollup.

Question 3

How does Semvec compare to Letta (formerly MemGPT)?

Accepted Answer

Letta runs as a server with its own database and uses LLM-driven page-in/page-out. Semvec is a pip-install library with a Rust core, no server required, and zero LLM calls on ingest. Letta is more flexible for free-form agentic memory; Semvec is faster, cheaper, and deterministic, with a fixed-size state and built-in multi-agent coordination via Cortex.

Question 4

Why is the Semvec core in Rust if the API is Python?

Accepted Answer

Determinism, throughput, and footprint. The semantic state update runs sub-millisecond on a single core. Bit-exact replay across runs comes for free — you can reproduce any session from a saved state. The Python layer stays familiar; the hot path stays cheap.

Question 5

Which embedders does Semvec work with?

Accepted Answer

Embedder-agnostic by design. Bring your own — sentence-transformers, OpenAI embeddings, ONNX models, anything that emits a fixed-dimension vector. The default examples use all-MiniLM-L6-v2 (384-d) on CPU; swap in whatever fits your latency budget.

Question 6

Does Semvec integrate with Claude Code and Cursor?

Accepted Answer

Yes. Semvec ships an MCP (Model Context Protocol) server. The SessionStart hook loads prior decisions, invariants, and error patterns into the prompt automatically. The PreCompact hook ingests the transcript and saves state before context compaction. The agent never has to be told to call save() — persistence is automatic across sessions.

Question 7

Does Semvec collect telemetry?

Accepted Answer

No. Anonymous pings were removed in v0.5.3. License verification happens locally via Ed25519-signed JWTs — no runtime phone-home. The library runs entirely on your infrastructure. Conversation content, embeddings, and state vectors never leave your machines.

Question 8

How does Semvec handle GDPR data deletion?

Accepted Answer

The compliance pack ships GDPR Article 17 (right to erasure) as async deletion with Ed25519-signed certificates, audit logging, HMAC signing, and RS256 JWTs. Verbatim cache uses Decimal precision so exact values (IBANs, dates, totals) are never lossy-rounded by cosine similarity.

Question 9

What if I want to remove Semvec later?

Accepted Answer

Zero lock-in. The chat proxy is a thin wrapper around your existing OpenAI-compatible client — drop the wrapper, your code keeps working. The library is a normal pip dependency; uninstall and the Python code stops importing it. State files in .semvec/ are JSON-shaped artefacts you can keep or delete.

Question 10

What does Semvec Cortex add for multi-agent workflows?

Accepted Answer

Cortex ships a checksum-validated shared state primitive across agents — no other dedicated memory system (mem0, Letta, LangMem, Zep, Cognee) ships a comparable primitive. Every StateVectorPacket carries a SHA-256 over the shared state; the receiver verifies before applying, so read-after-write drift becomes a verifiable error instead of a silent divergence. ConsensusEngine runs proposal voting at five levels (SIMPLE_MAJORITY, QUALIFIED_MAJORITY, UNANIMOUS, WEIGHTED_VOTE, ADAPTIVE_THRESHOLD) with quorum-aware finalisation. Per-agent attribution at retrieval surfaces cross-agent fact contradictions instead of returning one of the two at random. Deterministic replay across the entire agent fleet, within a release.

Question 11

Is Semvec only for healthcare?

Accepted Answer

No. Regulated Multi-Agent Memory is an architectural form, not a vertical product. The same five-component pattern — domain entity, specialist agents, shift boundary, audit trail, conflict resolution — applies identically to healthcare ward rounds, automotive pre-series-to-series engineering, insurance complex-claims handling, pharma cross-phase compound transfer, and legal multi-attorney case handover. Healthcare is our day-1 reference implementation because that is where our first lead landed; the other four are pattern-mapped, with automotive at the architecture-conversation stage. The Semvec + Cortex + Neo4j primitives are industry-agnostic.

Question 12

What is the INVESTIGATED edge in the Neo4j integration?

Accepted Answer

INVESTIGATED is a Neo4j relationship that records every agent retrieval and state update with 14+ canonical properties — drift_phase, drift_score, top_k_similarity, short_circuit, llm_call_id, agent_role, state_sha_before, state_sha_after, timestamp. It bridges the Semvec agent-session entity to your domain entities (Patient, Vehicle, Claim, Compound, Case File). One Cypher query answers 'what did this agent know at timestamp X and what did it do with that knowledge'. The reference repo demonstrates the pattern against a free Neo4j Aura instance — 15 minutes from zero to a working multi-agent demo.

Question 13

Can Semvec be deployed air-gapped?

Accepted Answer

Yes. Pre-built wheels run on your own infrastructure. License verification is purely local via Ed25519-signed JWTs — no runtime phone-home, ever. Offline-issued JWTs work in fully air-gapped environments. There is no init ping, no per-call event, no usage tracking. Versino PsiOmega GmbH is an EU vendor (Germany), so no US data-residency negotiation is required for EU procurement. Patent applications pending — EP 25 188 105, EP 26 160 795, US 19/269,195, US 19/550,466.

Claude Code that remembers your design decisions across sessions — without re-reading the diff.