v1.8.1 — MCP Light Memory

Full Changelog: v1.8.0...v1.8.1

v1.4.0 — retrieval improvements + temporal knowledge lifecycle

1.4.0

Section-aware chunking, read-only search, SimHash dedup, multilingual profiles, temporal knowledge lifecycle.

Task 5 — Section-aware chunking

• chunk_memory(): splits by Markdown headings, prefix with title/type/tags/scope
• Short memories get exactly 1 chunk; deterministic chunk IDs <memory_id>:<section-slug>:<ordinal>
• Config: retrieval.chunking.*; SQLite schema v3

Task 6 — Read-only search / usage tracking

• Search/context never mutate Markdown; usage lives in the SQLite usage store
• migrate-usage --dry-run/--apply [--strip] [--json] with timestamped backups
• index --rebuild preserves usage; content_hash excludes usage fields

Task 7 — SimHash deduplication

• Exact: SHA-256; Near: 64-bit SimHash (pure stdlib), Hamming <= 3
• remember/import: exact block, near warn, --force bypass; archived informational
• remember --json returns status + duplicate + separate conflict list

Task 8 — Multilingual PL/EN profile

• retrieval.profile: english-fast (default) | multilingual (intfloat/multilingual-e5-small, query:/passage: prefixes)
• Cache keys include model identity — profiles never share vectors
• Benchmark: multilingual > english-fast on PL group (R@1 12.5%→18.75%), no EN regression
• pack.py --profile english-fast|multilingual

Task 9 — Temporal, safe knowledge lifecycle

• Optional schema-2 frontmatter: confidence, valid_from, valid_to, supersedes, derived_from (schema-1 compatible)
• supersede <ref> --by <new>: closes validity window, links both directions; history never deleted
• search --at YYYY-MM-DD: validity-window filter; history stays retrievable; bad dates ignored
• timeline: sorts by effective validity
• context: read-only HISTORY & CONFLICTS section
• consolidate --dry-run [--json]: deterministic read-only report + plan for the agent; no deletion, no LLM summarization

Tests

• 115 unit tests OK (incl. tests/test_lifecycle.py)
• self_test.py: PASS