Recall is a function call,
not an API call.
Telys is the governed, embedded memory engine for AI agents — filtered vector search and temporal truth as a single function call inside your process. No daemon. No meter. No roundtrip. “Nearest where key = x” is one contiguous-slice scan, not a scatter across an index that ignores your keys.
0.013 msmeasured p50 — filtered top-10 over 1,000,000 vectors · recall 1.000 · in-processsee the conditions →
7/8CoIR code-retrieval tasks won vs reproducible BM25 — nDCG@10, identical mteb scoring pathsee the scorecard →
pipx install telys telys login # one sign-in → signed device license telys mcp install # Claude · Cursor · Codex
from telys import Telys db = Telys("./memory") col = db.create_collection("docs", dim=768, partition_by="tenant_id") col.add(vectors, ids=ids, metadata=metadata) hits = col.search(qvec, where={"tenant_id": "acme"}, top_k=10, explain=True)
Sign in once. Fully offline thereafter — the runtime is signed, licensed to your device, and never phones home at query time.
Not in someone else's cloud
The engine links into your process. Zero network roundtrips at query time — import the SDK, unplug the network, and queries still answer.
Time is the data model
Validity windows, supersession chains, provenance, and as_of() time-travel on MVCC snapshots. Ask your agent what it knew — and when it knew it.
License the engine, not the memories
Recall is free at runtime on every tier. No per-query, per-token, or per-retrieval line item — remembering is not a billable event.
The fast path is the default path
A key-scoped query resolves to a directory lookup plus one sequential scan of a contiguous block — recall 1.0 on the exact path. The speed is the layout, not a faster kernel: on the same contiguous subset our scan ties raw FAISS. Measured figures and their conditions live on the benchmarks page.
Filtered vector search as one contiguous memory operation.
Partition-key queries resolve to a directory lookup and one sequential scan of a contiguous block — with recall 1.0 on the exact path and an explain plan on every result. Off-key filters fall back honestly, and tell you so.
hits["explain"]["plan"]# on the partition key → "PartitionSliceExactF32" # one contiguous scan, recall 1.0# oversized partition → "PartitionIVFRerankF32" # IVF + exact rerank# off-key filter → "ScatterGatherExact" # honest fallback, and it says so
Powering Codna.
Codna — the code-intelligence product in the Thyn portfolio — links Telys as a library to serve scope-isolated code retrieval: structural code units partitioned by repo, service, and language, every scoped query one contiguous-slice scan with provenance and an explainable plan.
Read the methodology“No public "X× faster" claim ships until it is measured on the real workload.”
Telys engineering doctrine, D-28Built on foundations you already trust.
Apache Arrow in memory. Parquet segments on disk. FAISS at the ANN core. Hot loops run in Mojo SIMD kernels — a zero-Python hot path inside your Python process.
Stay in the loop.
Occasional release notes and roadmap updates — no noise. Ready to build now? Telys is Apache-2.0 and on PyPI: pipx install telys, then telys login (free for one developer).