lean-memory 0.1.3 — publish-readiness fixes

Publish-readiness release: an independent multi-team review of v0.1.2 found
three launch blockers on the canonical MCP first-run path plus packaging and
correctness majors — all fixed here.

Fixed

• MCP Registry install crashed on startup: server.json ran
  uvx --from lean-memory lean-memory-mcp, but mcp is an optional extra, so
  every registry install died with ModuleNotFoundError. The manifest now
  installs lean-memory[mcp,models,extract] (pinned by a manifest test).
• Model banner corrupted the MCP stdio stream: gliner2's from_pretrained
  prints a config banner to stdout — the JSON-RPC channel — on the first
  memory_add of the canonical install. All model lazy-loads (GLiNER2,
  SentenceTransformer, CrossEncoder) now route load-time chatter to stderr.
• Embedder swap bricked existing namespaces: reopening a DB created with a
  different embedder dimension (768-dim offline stub → 1024-dim Qwen after
  installing [models]) failed deep in retrieval with an opaque shape error;
  the store now refuses the mismatch at open with an actionable message.
• Uppercase FTS5 operator words crashed search: 'coffee AND tea' raised
  sqlite3.OperationalError through Memory.search and the memory_search
  tool. Terms are now quoted FTS5 string literals; the sparse arm degrades to
  no-hits on any residual syntax error.
• Functional-slot supersession left stale facts current: a replacement
  retired only the single most-similar fact, so a slot extended by an additive
  cue ("I also work at Globex.") kept two conflicting current employers. A
  replacement on a functional slot now retires every co-valid latest fact;
  multi-valued slots keep single-target retirement.
• High-similarity band ignored multi-valued slots: with real embedders,
  distinct co-valid values on a multi-valued slot (jazz/blues) embed at cosine
  0.6–0.95 and were silently superseded; multi-valued predicates now stay
  co-valid in every band (new resolver route: high_extends_additive).
  Predicate-scoped on purpose: the textual cue ("and"/"also") stays a
  low/mid-band signal so a conjunction-phrased replacement still supersedes.
• [llm] extra crashed every add() with Ollama stopped: Memory.add now
  catches TyperError and stub-types the escalated batch, as the typer
  contract always documented.
• Packaging: Apache-2.0 LICENSE added (repo, wheel, and sdist — the
  license was previously declared but its text shipped nowhere); the sdist is
  scoped to user-facing files (0.1.2 shipped internal strategy docs, agent
  instructions, and the bench harness to PyPI); the README hero GIF uses an
  absolute URL so the PyPI page renders it; the demo-agent flow is clone-based
  (the script was never in the wheel).

Added

• Schema-version stamp (PRAGMA user_version = 1) as the migration anchor for
  0.1.x namespace files (pre-stamp files upgrade in place; newer stamps are
  never downgraded).
• LM_FORCE_STUBS env var pins the offline stub backends in the MCP server
  (for tests/CI that must never load a model).
• Subprocess-level MCP stdio protocol test: handshake + real tool call, every
  stdout line must parse as JSON. CI matrix now covers Python 3.11/3.12.

lean-memory 0.1.0 — first public release

[0.1.0] - 2026-07-12

First public release. lean-memory is an embedded, local-first agent-memory
engine: one SQLite file per namespace, hybrid dense+sparse retrieval with
rerank, and ADD-only supersession queryable at any past point in time
(as_of). No server, no daemon, no mandatory cloud key.

Added

• MCP server exposing memory as three tools (memory_add, memory_search,
  memory_clear) for Claude Code, Claude Desktop, and other MCP clients.
  Canonical install pip install 'lean-memory[mcp,models,extract]'
  opportunistically upgrades each backend whose extra is present (real embedder
  • reranker via [models], GLiNER2 extraction via [extract]) and otherwise
    falls back to deterministic offline stubs. Two-minute quickstart with
    copy-paste Claude Code / Claude Desktop config and a demo GIF.
• Memory.search(now=...) — recency decay now anchors to a caller-supplied
  timestamp, so the 0.2 recency term is no longer dead on historical corpora.
• Point-in-time queries via as_of (epoch ms) with is_latest_only=False.
• CI + release workflows (GitHub Actions): offline test matrix on
  ubuntu/macOS × Python 3.10/3.13, plus build-and-publish to PyPI on v* tag
  via Trusted Publishing.
• PyPI metadata: keywords, classifiers, and project URLs.

Changed

• Default embedder is now the ungated Qwen3-Embedding-0.6B (was a gated
  Gemma model that broke the [models] first run). Reranker default is
  Ettin-32M; both are pinned ungated and covered by regression tests.
• Escalation engine recalibrated on real conversational turns. Endpoint-
  scoped coreference/ellipsis detection replaces the whole-text pronoun scan
  (coreference escalations dropped from 65.6% to effectively nil on real
  turns), and the prior_entity trigger was retired (subject re-mention is
  normal discourse, measured at 52.8% of candidates). At the re-frozen
  (typing_threshold=0.4, conf_threshold=0.4) operating point, escalation on
  the real LongMemEval probe is 14.6% (was ~96% pre-fix), with the residual
  being irreducible inferential-edge (derives) escalations. BET-2 three-gate
  revalidation PASSes at this operating point.
• Extraction granularity calibrated — GLiNER candidate threshold set to
  0.4, cutting the extractor from ~8 facts/turn to ~3.7 so fact_text reads as
  facts rather than whole utterances.
• MCP server loads models lazily (first tool call rather than import) so a
  cold-cache spawn answers the MCP handshake immediately instead of blocking on
  a model download; search output is deduplicated.

Fixed

• Sparse BM25 retrieval arm now honors the as_of interval predicate.
• Known-entities handed to the router/typer are capped at the 100 most recent.

Install: pip install lean-memory · MCP quickstart: see the README

🤖 Generated with Claude Code