[FEATURE] Native cognitive memory — NEXO Brain achieves F1 0.588 on LoCoMo (+55% vs GPT-4)

The Problem

Claude Code starts every session from scratch. CLAUDE.md files are manual, static, and grow without bound.

The Solution We Built

NEXO Brain is an open-source MCP memory server (MIT) that implements the Atkinson-Shiffrin memory model:

• Automatic ingestion from conversations
• Semantic retrieval (768-dim embeddings + BM25 hybrid search)
• Cross-encoder reranking
• Multi-query decomposition for complex questions
• Adaptive Ebbinghaus decay (unique memories protected)
• Dream cycles for overnight consolidation

Benchmark Proof

Tested on LoCoMo (ACL 2024, peer-reviewed):

| System | F1 | Hardware |
|---|---|---|
| NEXO Brain v0.5.0 | 0.588 | CPU only |
| GPT-4 (128K full context) | 0.379 | GPU cloud |
| Gemini Pro 1.0 | 0.313 | GPU cloud |

+55% vs GPT-4 on long-term conversation recall. Running on a MacBook CPU.

Why Native?

As an MCP server, there's 50-200ms overhead per tool call. Native integration would mean:

• <1ms latency
• Zero user setup (built-in)
• Automatic context pre-fetch (no manual tool calls)
• 90-95% reduction in context tokens (selective retrieval vs full CLAUDE.md)

Traction

• 949 npm downloads on day 1
• 97+ MCP tools across 17 categories
• MIT license, ready to integrate

npm: https://www.npmjs.com/package/nexo-brain
Repo: https://github.com/wazionapps/nexo
Proposal: https://github.com/wazionapps/nexo/blob/main/pitch/anthropic-integration-proposal.md