Memory Bootstrap

This is the single skill every other skill in this system depends on. If a session doesn't start by reading project memory, nothing downstream (recall, checkpointing, handoff) has anything to build on — this is the entry point of the whole architecture described in ARCHITECTURE.md of the agent-memory-protocol repo.

What to read, in order, at session start

1. .memory/BOOTSTRAP.md — the compact (<150 line) orientation: what this project is, current architecture summary, active conventions, non-negotiables. This is the equivalent of a coding agent's AGENTS.md/CLAUDE.md, scoped specifically to memory-system orientation.
2. .memory/HANDOFF.md — the most recent explicit handoff note, if one exists (written by agent-handoff when a prior session deliberately switched tools/agents).
3. The most recent file in .memory/checkpoints/ — sorted by timestamp, this is the actual last-known state, which matters most specifically because a session might have ended without a clean handoff (credit exhaustion, crash) — see session-checkpoint's rationale for why this file, not HANDOFF.md, is the authoritative "where things actually stood."
4. Any .memory/tasks/*.md files marked active — current in-flight work, so the new session doesn't have to be told what's being worked on.

What this produces

A working understanding of: what the project is, what was being worked on, what decisions have already been made (don't re-litigate them without reason — check .memory/decisions/ before proposing something that contradicts an existing decision), and what the immediate next step was supposed to be.

The failure mode this prevents

A new agent starting cold — asking the user to re-explain the project, re-stating already-made decisions as open questions, or worse, silently contradicting a documented decision because it was never read. This is exactly the "no manual reconstruction, no rebuilding context" requirement — it only works if this skill actually fires automatically and actually reads the files, not if it's available but skipped.

Practical trigger discipline

This should fire before the first substantive response in any session where .memory/ exists in the project root — not after the user asks for it. If .memory/ doesn't exist yet, this is also the moment to note that and suggest initializing it (see the agent-memory-protocol repo's setup instructions), rather than silently proceeding without any memory system in place.

What NOT to do

Don't re-read the entire episode history (.memory/episodes/) at bootstrap — that's expensive and mostly unnecessary; memory-recall handles targeted retrieval of specific past episodes when something in the current task actually calls for it. Bootstrap is a fast, cheap orientation pass, not an exhaustive history read.