honcho — Wednesday’s shared long-term memory

Category: mcp Where it works: Claude Code · Telegram (Hermes uses it natively for memory) — both Source: _infra/mcp/servers/honcho/server.py → Honcho server on Jarvis (192.168.0.100:8005), workspace wednesday, peer sabin

What it is

The cross-runtime memory layer. It accumulates a peer card (durable facts about you, derived from conversation over time) + stores sessions/messages, so any Wednesday runtime can recall who you are and what was decided — not start fresh. This is what makes Telegram-Wednesday “remember where we left off.”

How to trigger it

You rarely call it by name — Wednesday queries it for you. Natural language that pulls it: “what do you know about me?”, “what did we decide about X?”, “have we discussed Y?”, “remember that Z”. On Hermes it’s automatic (memory provider).

How to use it

Ask Wednesday a memory question and she runs the right honcho tool. To deliberately seed a fact, say “remember that ” → honcho_add_observation.

Tools

toolwhat it doesexample
honcho_get_peer_cardread the accumulated key-facts about a peer”what do you know about me?”
honcho_chatdialectic Q&A — reasons over derived memory (reasoning: minimal→max)“what’s my stance on background Claude?”
honcho_searchsemantic search across all stored sessions”find when we discussed split-DNS”
honcho_get_messagesread messages in a specific session”show the messages from session X”
honcho_list_peerslist known peers(discovery)
honcho_list_sessionslist all sessions in the workspace(discovery / audit)
honcho_add_observationwrite a one-off fact about a peer”remember that I prefer Telegram for mobile”

Notes / limits

The deriver/dialectic LLM runs on Grok (via the Hermes proxy) — cheap, not Claude. The peer card is derived (grows as you talk; a fresh fact may take a bit to surface). Embeddings off by default. Honcho server lives on Jarvis — if it’s down, recall degrades (Wednesday still works, just without long-term memory).