AliceLJY/recallnest

RecallNest v2.0.0 — Memory Operating System

124 commits since v1.4.0. RecallNest evolved from a memory store into a full memory operating system.

Major Features

Memory Intelligence

• Session Distiller — 3-layer conversation compression: microcompact → LLM structured summary → knowledge extraction to durable memory
• Conversation Import — bring your history from Claude Code, Claude.ai, ChatGPT, Slack, or plaintext (auto-detected)
• Topic Tags — intra-scope topic partitioning with 15 auto-detected topics, filterable in search
• Dream Pipeline — 4-stage offline memory consolidation (clustering, merging, pruning, pattern extraction)
• Skill Memory — store, retrieve, and auto-promote executable skills from recurring patterns

Retrieval Evolution

• Multi-hop Retrieval — entity-driven two-round search for cross-session evidence expansion
• Knowledge Graph — KG triple extraction + PPR graph traversal + interactive D3.js visualization with semantic bridges
• Multi-vector Search — L0/L1/L2 dynamic folding with 6-channel hybrid fusion
• Recall Governor — auto-recall governance layer with circuit breaker
• 4 Retrieval Profiles — default, writing, debug, fact-check (tuned for different tasks)

Memory Lifecycle

• Memory Evolution Engine — supersede chains, Weibull decay, LLM importance scoring, auto-archival
• Admission Control — write-time gating: noise filter, importance floor, dedup, rate limiting
• Memory Lint — content quality checker with health score (contradictions, duplicates, stale, orphans)
• Context Collapse — dynamic folding view with staleness hints

Multilingual (babel-memory integration)

• BM25 pre-tokenization for CJK languages (jieba / kuromoji / Snowball)
• Language-aware KG extraction prompts (EN/CJK)
• Language-aware session summary prompts (EN/CJK)
• Auto language detection stored in metadata

Product Experience

• Web UI v2 — Dashboard (stats, category distribution, health, growth) + Search Workbench (topic filter, Skills tab)
• Claude Code Plugin — one-command install via marketplace
• Ultra-Light Wake-up — resume_context(mode='light') returns <300 tokens
• 40 MCP tools in 3 tiers (core / advanced / governance)
• 21 HTTP API endpoints

Metrics

| Metric | v1.4.0 | v2.0.0 |
|--------|--------|--------|
| Source files | ~60 | 108 |
| Lines of code | ~15K | 31,800+ |
| Tests | ~600 | 1,213 |
| MCP tools | 20 | 40 |
| HTTP endpoints | 12 | 21 |
| Retrieval channels | 2 (vector + BM25) | 6 (vector + BM25 + L0 + L1 + L2 + KG) |

Benchmark: LongMemEval (ICLR 2025)

| | RecallNest | Vector-only | Delta |
|---|---|---|---|
| Overall Accuracy | 29.6% | 24.2% | +5.4pp |
| User Facts | 64.3% | 52.9% | +11.4pp |
| Abstention Rate | 55.6% | 67.8% | -12.2pp |

Full changelog: https://github.com/AliceLJY/recallnest/compare/v1.4.0...v2.0.0