Changelog

All notable changes to agentmem will be documented in this file.

[0.2.0] - 2026-03-30

Truth governance engine — memory lifecycle states: hypothesis -> active -> validated -> deprecated -> superseded
Provenance tracking — source_path, source_section, source_hash fields on every memory
Conflict detection — finds contradictions between active memories, separates duplicates from real contradictions, sentence-level negation matching
Staleness detection — flags memories by age, missing source files, or hash drift (source content changed)
Health check — scores memory system 0-100 based on conflicts, stale %, orphaned references, validated count
Trust-ranked recall — validated memories rank above active, which rank above hypothesis; provenance-weighted (canonical > unprovenanced)
Lifecycle methods — promote(), deprecate(), supersede() on Memory class
Session governance — save_session() now properly supersedes previous sessions
MCP governance tools — promote_memory, deprecate_memory, supersede_memory, memory_health, memory_conflicts
CLI governance commands — health, conflicts, stale, promote, deprecate
28 governance tests — lifecycle, conflicts, staleness, hashing, search trust ranking
Schema v2 migration (backward-compatible from v1)

Search no longer bumps access_count on returned candidates (fixes self-reinforcing popularity bias)
Deprecated and superseded memories excluded from search results entirely
Recall ranking reweighted: FTS 25%, trust status 20%, provenance 20%, recency 15%, frequency 10%, confidence 10%
README rewritten around "governed memory for long-lived coding agents"
Package description updated

Resurrected sections: if a canonical section is removed (deprecated) then restored, sync reactivates it
Source priority: provenance-tracked canonical memories now rank above unprovenanced imports
Session lifecycle: old sessions properly superseded, no more orphaned chains

Core Memory class with CRUD operations
FTS5 full-text search with porter stemming
Context-budgeted recall() with token limits
Seven memory types: setting, bug, decision, procedure, context, feedback, session
Project scoping (multiple agents, one DB)
Session persistence (save_session, load_session)
Markdown importer with frontmatter and type inference
CLI with full CRUD, search, recall, import, stats
MCP server with 8 tools
SQLite + WAL mode, thread-safe
31 tests across core, search, and importer
Published to PyPI as quilmem