Changelog · pre-1.0

What we've shipped.

Locamem is pre-1.0. We ship dated changelog entries instead of marketing version numbers — every line below maps to a real change in the engine, the MCP server, or the site. The local engine is and stays MIT-licensed and free; nothing here meters memory or charges per recall.

June 13, 2026
Brand + site

The Locamem brand and a new site

  • Launched the Locamem brand and made locamem.com the canonical domain.
  • New editorial, trust-first marketing site: home page plus dedicated legal, healthcare, and finance vertical pages.
  • Published a /benchmarks page with the honest numbers — LongMemEval-S session recall (99.4% recall@10) and the reader-limited ~58% end-to-end QA ceiling, not cherry-picked wins.
  • Stood up the docs site at docs.locamem.com.
June 13, 2026
P0 — MCP + install

Two-transport MCP server and one-line install

  • Shipped an MCP server on the official SDK with two transports: stdio and streamable-HTTP.
  • Exposed 10 MCP tools for storing, recalling, and inspecting memory over either transport.
  • Added a one-line installer: curl -fsSL https://locamem.com/install | bash.
  • Installer auto-detects your MCP client and wires up the config for you.
May 29, 2026
v6

Deterministic solvers + air-gapped vertical demos

  • Added deterministic solvers to the recall path.
  • Ran an honest end-to-end ceiling investigation — measuring where the reader, not retrieval, caps QA accuracy.
  • Built air-gapped demos for legal, healthcare, and finance with live in-browser recall and zero network calls.
  • Switched the site to clean URLs.
April 8, 2026
Reasoning + packaging

Chain-of-thought answers and graph-based recall

  • Added type-aware chain-of-thought answer generation.
  • Added graph-based reasoning to the recall pipeline.
  • Published the honest results of the day: 93.6% retrieval, 52% QA.
  • Started shipping the package on PyPI.
Initial
Core engine

Single-file memory with auditable recall

  • The whole memory is one local SQLite file — no server, no service to run.
  • Retrieval combines SimHash (LSH), FTS5 full-text, and optional on-device embeddings.
  • Tracks temporal validity and an entity lattice, with salience scoring over stored facts.
  • Every result returns a per-facet score breakdown — recall is auditable, not a black box.

Following along? Watch the repo at TeamWilcoe/locamem for the full commit history.