Introduces the zosma-mem package — a tag-based, salience-driven memory system for OpenZosma agent sessions.

[Shanvit7]

Summary

Core memory engine

• Entity store with stable deterministic IDs (SHA-based content hash)
• Co-access graph for boosting related entities during retrieval
• Two-pass retrieval: tag overlap + salience scoring, then co-access re-rank
• Ingestion pipeline for extracting and storing facts from agent turns
• Salience decay + GC to prune low-value entities over time
• Reinforcement signals (used / ignored / influenced_decision)

Bridge

• createMemoryBridge wires the engine to the pi-mono session via a
  loadContext / ingestFacts / recordUsage interface
• Memory directory scoped per agent config
  (workspace/agents/<agentConfigId>/memory/)

Eval pipeline

• SQuAD v1.1 validation set used as a benchmark (fetched from HF)
• Content-based precision metric: scores 1 if retrieved context contains
  the expected answer substring — more honest than ID-based precision for
  a topK retrieval system
• CLI: zosma-mem-eval run with configurable case count

Baseline scores (100 cases, topK=8)

Metric	Value
Recall	72%
Precision	77%*
F1	72%

*ID-based precision ceiling is 1/8 = 12.5% by construction.
Content-based precision (new metric) reflects real usefulness.

Related issue

Checklist

• pnpm run check passes
• pnpm run lint passes
• Tests added or updated (if applicable)
• Documentation updated (if applicable)