rohitg00/agentmemory v0.7.6

v0.7.6 — cross-encoder reranking, working memory, skill extraction

on GitHub

New Features

Cross-encoder reranking (#90)

Opt-in rerank step after hybrid search via RERANK_ENABLED=true. Uses @xenova/transformers MiniLM-L-6-v2 cross-encoder on top-20 candidates. Graceful fallback when model unavailable.

Working memory / context window (#58)

Core memory system with pinned facts (always included in context) + automatic paging of archival memories. 5 new functions: mem::core-add, mem::core-remove, mem::core-list, mem::working-context, mem::auto-page.

Skill extraction (#57)

Auto-generate reusable procedural skill documents from completed sessions via LLM. Idempotent reinforcement. mem::skill-match queries relevant skills. 3 new functions.

Bug Fixes

• Plugin install validation (#94): Fixed plugin.json — author must be object not string, removed hooks field (auto-loaded by convention), skills as explicit path array

Improvements

• Rerank window decoupled from limit (fixed 20-item cross-encoder window)
• Parallel KV fetches in skill extraction
• continue instead of break for oversized core entries
• Access-counter writes buffered with Promise.allSettled
• RERANK_ENABLED parsed with === "true"
• All state-changing operations audited
• Session status guard for skill extraction
• Skill fingerprint covers title + trigger + steps

Closes #90, #58, #57, #94

Full Changelog: v0.7.5...v0.7.6