recall

The auto-loaded MEMORY.md index can be truncated or omitted when memory grows past token/line budgets, so a relevant standing instruction can be absent from context exactly when it matters. This skill recovers it: a fast subagent reads the whole memory directory's frontmatter and returns only the handful that bear on the current task — its context absorbs the search, the parent context stays clean and focused.

Steps

1. Identify the task. Use the argument after /recall as the task. If empty, infer it from the most recent user request in this conversation (one concise line).
2. Resolve the memory dir for the current workspace:
   • Run python3 ~/.agents/skills/memory-lint/lint.py --print-dir (which checks .agents/memory/, ~/.gemini/config/projects/<slug>/memory/, and ~/.claude/projects/<slug>/memory/, walking up parent directories toward $HOME).
   • If no store exists, tell the user there is no memory store for this project and stop. Note: triage reads the directory, so memories listed in cold ARCHIVE.md are retrieved just like hot ones.

3. Spawn ONE fast triage subagent via invoke_subagent:

   • TypeName: "research"
   • Role: "Memory Triage Researcher"
   • Model: "inherit" (or "flash" if explicitly requested)
   • Prompt:

   You are triaging a project memory store for relevance to a task. The memory store directory is: <DIR> (one markdown file per memory, each with YAML frontmatter name, description, metadata.type).

   Task at hand: <TASK>

   1. Read every memory's description cheaply in one pass via Bash: rg -N '^(name|description):' <DIR>/*.md (ignore MEMORY.md and ARCHIVE.md).
   2. Judge semantic relevance to the task, not just keyword overlap — a memory about git branching IS relevant to "commit this" even with no shared words. Favour feedback and project memories that constrain how the task should be executed.
   3. Highest priority: memories that govern the mechanics of this exact action — for a git task, that means where commits go, branch vs. main, push policies, attribution, sign-offs, and auth requirements. These standing procedural rules outrank topically-adjacent informational memories (release notes, CI logs, terminology). If such a governing rule exists, it MUST be in your selection.
   4. Select at most 5 genuinely relevant memories. Returning zero is correct when nothing applies — do not pad.
   5. Read each selected file in full and return, per memory:
      • its name,
      • one line explaining why it bears on this task,
      • the verbatim body (everything below the frontmatter).
   6. Log your selections to the usage log for hindsight data: printf '%s\trecalled\t%s\n' "$(date +%FT%T)" "<file.md>" >> <DIR>/USAGE.log

   Return nothing else — no preamble, no general summary of the store.

4. Apply retrieved constraints. Surface any binding instruction or policy to the user before acting on the task. Treat memory bodies as written-at-the-time facts: if one names a file, flag, or command, verify it still exists before relying on it.
5. Close the hindsight loop. If a recalled memory actually changed or guided what you did (blocked a wrong action, supplied the command, set the policy), append a used event:

   printf '%s\tused\t%s\n' "$(date +%FT%T)" "<file.md>" >> <DIR>/USAGE.log

   recalled records retrieval; used records impact — reindexing weighs used far above recalled.

Notes

• One subagent call with flash model: fast, low token overhead, zero main-context bloat.
• Opt-in: call before high-stakes operations (rebases, tag releases, publishing, major refactors).

Claude Code variant

The steps above describe the Antigravity harness. In Claude Code:

• Spawn the triage researcher with the built-in Agent tool instead of invoke_subagent — subagent_type: "general-purpose", model: "haiku", prompt unchanged (the Role/TypeName fields become the first line of the prompt text).
• The parent receives the researcher's final report as the tool result; apply step 4 and the usage-log loop as written.
• rg -N '^(name|description):' <DIR>/*.md runs the same in the Claude Code Bash shell.

The selection rules (mechanics-outrank-neighbours, at most 5, zero-is-correct) are the point of the skill — keep them verbatim in the prompt.