titen-memory 0.6.0

Complete changelog

Stable channel · npm latest · pre-1.0, the public API is not yet stable

npm install -g titen-memory@0.6.0

Install guide · Release page · npm package

First release with a published head-to-head against a comparator on an externally authored corpus, and the first where a measured result caused work to be deleted rather than shipped.

Measured against self-hosted Mem0: parity, not superiority

• Mr.TyDi Indonesian, 25 queries over 100 documents, 10 repeats, one shared embedding endpoint, against mem0ai 2.0.13 in library mode.

Point estimates favour Titen on every metric. That is not a win. Both systems are deterministic on this harness, so disjoint across-repeat ranges prove nothing; the uncertainty that governs a between-system claim is query sampling, and paired sign tests give p ≥ 0.500 on every metric. The whole margin is two queries — 20 of 25 rank-1 hits against 18 — and Titen loses one query to both Mem0 configurations.

What the run does support: retrieval parity, ~1.9× lower query latency (~400× without vector search), ~9.6× faster ingest, and deterministic ranking when vector search is enabled.

Deleted rather than shipped

• An adversarial review rejected the first retrieval change. Isolated at 10 repeats, the score-formula rewrite earned nothing the tie-break alone does not, regressed the FTS-only lane, and coupled the lexical/semantic balance to one provider’s cosine distribution through a fitted constant. It was removed; what shipped is four lines in the comparator. Issue #227 stays open and #226 is narrowed rather than closed, in the changelog and in the debt ledger.

Corrected in the record

• Figures published earlier the same day for 0.5.7 — recall@1 0.680, MRR@10 0.807, query p50 170 ms — were the top of their ranges over five repeats, not medians. The true 0.5.7 median is 0.62. They should not be requoted.

Newly measured, and uncomfortable

• The service uses one core at every concurrency level, so a single client saturates it at 10,000 claims, and retrieval quality falls with corpus size (recall@1 1.00 to 0.81 to 0.49 across 10³, 10⁴, 10⁵ on a synthetic corpus). Cold start stays around 53 ms and resident memory grows 2.9% across a hundredfold corpus. The data-usable upgrade floor is 0.2.0: a 0.1.x store migrates without error and then answers 404 on every claim.

Added

• An optional top_k on context compile, a standard-library-only Python client, a container image, and the release-bound retrieval harness that the evaluation spec had been waiting for.