Long-Term Memory and Knowledge Graph AI Foundation

Problem It Solves

Agent sessions often lack stable memory, so users re-explain context; memory plugins also store data separately, making cross-skill reuse hard. As memory grows, duplicates, conflicts, bloat, and noisy retrieval become likely. zwjh-skill turns long-term memory, knowledge graphs, retrieval, compression, and backup into a local, callable foundation rather than another isolated plugin.

How It Works

• Long-term memory: distills knowledge from conversations, files, and daily logs while preserving prior logs and idempotently backfilling entries.
• Knowledge graph: extracts entities such as people, projects, tasks, events, and documents, with relations and reusable interfaces like add-entity and relate.
• Automatic deposition: merges exact or near-duplicate items and marks competing values for the same entity and predicate as superseded.
• Local retrieval: uses TF-IDF Chinese vector cosine similarity and date-range timeline search, with no API key and no network dependency.
• Health and backup: audits stale memory, orphan entities, conflicts, and DB size; supports snapshots, compression, restore, and optional local or Baidu Netdisk backup.

Boundaries

It fits agent workflows that need local-first, cross-session memory and controlled governance. Graph quality depends on input text; conflict handling preserves history but does not automatically verify factual truth. Baidu Netdisk is pluggable, off by default, and does not trigger network access automatically.