ambrogio-memory

The lab’s memory service: one Go binary with an embedded SQLite file that owns the present state, the log, the debt register, the documents and the closed cycles. Agents read and write it through an HTTP API; the markdown tree is not the memory any more.

• Repository: github.com/tazzo/ambrogio-memory (private, main)
• Runs on: CT106, systemd user unit ambrogio-memory.service, caged with Landlock
• Data: /opt/ambrogio-memory/data/memory.db (the one thing that cannot be lost)
• API: http://192.168.1.206:7810 (LAN and tailnet)
• Skill: /workspace/SKILLS/memory/SKILL.md — the client contract, served by GET /api/v1/skill
• Backup: the whole memory as markdown under SKILLS/memory/backup/, committed and pushed every 30 minutes by the service itself
• Deployed commit: 66ef795 (2026-10-21)

What it is not

It spawns no agents and no processes: the only thing that thinks is one internal agent that calls a model to write the account of a closed cycle. There are no queues, no tasks, no workers, no second door, no web page and no secrets in the database.

Canonical Starting Pages

• topics/ambrogio-memory-contract — the model: rows, boot bundle, compaction, the internal agent, backup and restore.
• details/ambrogio-memory-verbs-detail — every verb, its parameters and its refusals.

Relationships

• entities/agents-ctx — the skills workspace, which carries the memory contract and the backup tree.
• entities/mnemosyne-mcp-server — the semantic retrieval layer; it is not the record.