Released · PyPI verified
MADB 0.2.14

Memory that recalls, corrects, anticipates and knows when to abstain.

This release extends 0.2.13 current-truth recall into a controlled episodic-memory lifecycle for autonomous agents.

Cognitive lifecycle

What the release adds

  • Typed epistemic memory: separate observations, hypotheses, directives and outcomes.
  • Replay consolidation: deterministic selection with immutable receipts.
  • Controlled reconsolidation: correct authority while preserving history.
  • Adaptive salience: reversible suppression, quarantine and reactivation.
  • Prospective memory: time and context cues with exactly-once claims.
  • Executive context: bounded, diverse working-memory admission.
  • Metamemory: coverage, conflict and explicit abstention signals.
Temporal knowledge

Observed time and available time become first-class.

The release adds indexed, tenant-scoped temporal queries so an agent can distinguish when an episode described the world from when that knowledge became available. This preserves point-in-time truth without exposing or coupling MADB to any application-specific implementation.

Release status

Released from verified commit ca9010c after cross-platform GitHub CI; engine, MCP server and Anthropic SDK packages are published on PyPI.

Review executable evidence