Smaller prompts per reply
Around 2,000 tokens of context per reply under benchmark conditions instead of the entire history.
LOCOMO · gpt-4o-mini reader · T=0 · vs full-contextMemory for AI agents · made in Munich
Every conversation, one memory.
Semvec is a Python library and proxy that gives your AI agents a permanent, fixed-size memory. Everything your agents hear converges into one compact core, so they remember every conversation and your token bill stays flat. pip install semvec
Scroll
Without Semvec
Every conversation bends your budget.
A normal agent re-reads its entire history for every single reply. Each running conversation becomes a gravity well in your budget: costs climb, latency grows, context windows overflow, pilots stall.
Reply 50 re-reads0tokens
With Semvec
One core. 4‑5k tokens. At turn 10 and 10,000.
Every exchange folds into a memory of constant size. Storing costs zero AI calls, runs in under a millisecond on your own hardware, and is deterministic: the same inputs always produce the same memory. What never grows can never blow up your bill.
The whole idea in one picture
The conversation keeps flowing. The memory never grows. That is the entire trick, and everything below is what it does for your business.
The product
Meet the instrument.
A Rust core with a Python surface. It sits between your application and any OpenAI-compatible AI provider and does one job with precision: remember, at fixed cost.
Inside
Every part has a job.
Proxy shell · Memory core · Zero-cost intake · Savings meter · Audit seal
Integration
Three moves. No rebuild.
01
Point your AI at the Semvec proxy
Change one base URL, keep your SDK and code. One command installs it: pip install semvec, prebuilt for Linux, macOS, and Windows, Python 3.10 to 3.14.
02
Every exchange folds into the core
No stockpiling, no nightly cleanups, no extra AI calls. Deterministic math on your own hardware, done in under a millisecond.
03
Every reply ships with proof of savings
Each response reports what it cost with Semvec next to what the full history would have cost. Measured on your own traffic, not a vendor slide.
For your engineers: the full technical documentation
Measured, not promised
0
AI calls needed to store a memory
Storing is plain math. The best-known competitor makes a generative AI call for every exchange it stores, about 5,000 across the standard benchmark.
LOCOMO · mem0 one call per stored turn
~87%
17x
Faster through the full benchmark
The complete LOCOMO suite in ~60 minutes versus ~16.7 hours with mem0 on the same machine.
Full-suite wall-clock · GPU reference platform · hybrid retrieval on
Replay
Auditable by construction
Deterministic updates: any past memory state can be replayed and inspected. Your auditors will ask for exactly this.
Bit-identical replay within a release · GDPR deletion certificatesThe honest trade: mem0 scores about 6 points higher on the LOCOMO answer-quality judge (0.669 published, 0.675 measured, vs our 0.605); we publish both. If that margin rules your case, pick mem0. If cost, speed, and auditability rule it, pick Semvec.
Semvec Cortex
Ten agents. One memory.
Teams of specialist agents share one memory: handovers sealed with checksums, decisions taken by configurable consensus votes, and a complete answer to the question every auditor eventually asks: who knew what, when.
HealthcareAutomotiveInsurancePharmaLegal
Built to be bought
Give your agents a memory.
- EUVersino PsiOmega GmbH, Munich. Self-hosted, air-gapped, or Versino-managed in the EU.
- GDPRRight-to-erasure built in, with signed deletion certificates and an append-only audit store.
- OFFLINENo telemetry. Licence keys verify locally; three tiers from free evaluation to enterprise SLA.
- IPFour patent applications pending (EP and US).
- PRESSFeatured on the Neo4j Developer Blog.
Typical reply within one working day.