iiinova — where your digital self keeps evolving

ogham-mcp @ogham-mcp.bsky.social · 29/03/2026

Anthropic shipped 3 tools for async agent work this week. But the most interesting part of Nate B Jones' breakdown isn't the tools. It's what he says about memory.

100

ogham-mcp @ogham-mcp.bsky.social · 25/03/2026

Claude Code now writes notes to itself between sessions. Auto-memory + auto-dream consolidates them every 24 hours in the background. It's good. But open the same project in Kiro or Cursor and nothing transfers. We built Ogham for that gap. ogham-mcp.dev/blog/claude-auto-memory/

ogham-mcp.dev

Claude now remembers things. Here's what it doesn't remember.

Claude Code's auto-memory and auto-dream are good steps for single-client persistence. But what happens when you switch editors?

100

ogham-mcp @ogham-mcp.bsky.social · 25/03/2026

Tested Gemini Embedding 2 at four MRL truncation points: 3072, 1536, 768, 512 dims. The scores went UP as I compressed. 512 dims outperformed 3072 on every single query. Rankings stayed identical. Full results: ogham-mcp.dev/blog/gemini-embeddings-multimodal/

ogham-mcp.dev

Gemini Embedding 2 at 512 dims - does multimodal survive compression?

We tested MRL truncation from 3072 to 512 dimensions on cross-modal retrieval. The results surprised us.

100

ogham-mcp @ogham-mcp.bsky.social · 24/03/2026

Ran 400 memory retrieval questions BEAM benchmark against pgvector. No knowledge graph, no reranker, no LLM at search time. Our retrieval found more gold evidence than the best LLM (Llama-4 + LIGHT) could use. In 9 out of 10 categories. #ai #memory Full results: ogham-mcp.dev/blog/beam-benchmarks/

ogham-mcp.dev

101

ogham-mcp @ogham-mcp.bsky.social · 24/03/2026

I built an open-source MCP server that gives AI agents shared memory across clients. Store something in Claude Code, search for it from Opencode, Cursor. No LLM needed, just Postgres + embeddings model. MIT licensed. #opensource #ai #memory github.com/ogham-mcp/ogham-mcp

350