On the road to 0.4.0 — single-writer daemon, agent attribution, and the ND-Mem MCP App

[jmeyer1980]

Over the last few days we've landed a batch of features and fixes in the release candidates that take us from 0.3.9 toward 0.4.0. Here's what's new.

🧠 One brain, one writer: the single-writer daemon

The biggest change under the hood. Previously, every MCP client spun up its own server process, and multiple writers could silently overwrite each other's memories (last-writer-wins). Now a single daemon owns the memory store:

• --daemon mode — the sole process that opens memories.json, serving MCP over Streamable HTTP on 127.0.0.1:3838
• stdio proxy (the new default) — ensures the daemon is running and forwards JSON-RPC; it never opens the store itself
• standalone mode — the previous behavior, kept for tests, CI, and the inspector

No matter how many editors, agents, or apps you have open, there's exactly one writer. The data-loss failure mode is gone, and it's covered by a regression test.

🏷️ Agent attribution — know who wrote what

Memories now carry reliable agent identity, with a clock-in/clock-out model:

• Every MCP client gets a persistent per-connection session, and the daemon auto-binds agent_id from the client's initialize info. No more passing agent_id on every store_memory / update_memory call — attribution just works.
• server_handshake accepts an optional agent_id to override the auto-bound identity, and reports the session's current binding.
• Identity clocks out automatically: on a kind:handoff write, on close_task, or after an idle timeout (default 30 minutes, configurable via NEURODIVERGENT_MEMORY_SESSION_IDLE_MS).
• Clients that never handshake (the web-app bridge, bare HTTP calls) keep the exact prior stateless behavior — nothing breaks.

This also makes tool use leaner: fewer parameters on every call, and the bridge no longer spawns its own MCP child — writes route straight to the shared daemon.

🖥️ The ND-Mem MCP App

The web app (served by the local bridge) is now a proper front door to your memory store. It reads memories.json live, listens for changes, and routes every write through MCP — so the daemon stays the single writer even when you're editing from the browser.

Browse and read your memories with project auto-discovery, district carousels, search and filters, and a selected-card panel showing metadata, tags, and MCP routing details:

📸 [Screenshot 1 — main view]

Edit any memory in place. The edit modal routes to the bridge, which forwards the change to the MCP update_memory tool — only the fields you touched change:

📸 [Screenshot 2 — edit memory window]

Write new memories from the browser with the New Card form — title, district, project, visibility, intensity, content, and tags, posted through the bridge to store_memory:

📸 [Screenshot 3 — new memory window]

And a favorite: rename projects, with smart merging. Renaming to a name that already exists (matched case-insensitively, treating - and _ alike) merges into the existing project's canonical spelling — after asking you to confirm. A "did you mean…?" guard catches near-miss names before they become accidental duplicates.

🔧 Hardening

• tags are now validated as an array of strings at every write boundary, so malformed payloads can't crash synthesis tooling downstream
• The stdio proxy recovers cleanly from idle-timed-out sessions instead of hanging
• The bridge resolves its assets from its own file location, so it works when launched from any directory
• Daemon tool errors now surface as failed requests in the web app instead of disappearing silently

Try it

npm install neurodivergent-memory@rc

Feedback, bug reports, and weird edge cases welcome — that's what the RC line is for. 🚀