auto-skill-build-memsearch-codex-memory

Use memsearch as an on-demand semantic index for project markdown memory (search → expand → optional deep drill). Use when the user asks to recall past decisions, prior debugging, or “have we seen thi

SKILL.md

GitHub

memsearch + Codex memory recall (on-demand)

Generated by: Codex Supercharge maintenance automation.

This skill is for manual recall using the memsearch CLI. It does not install hooks or run upstream installers.

Reference notes: references/memsearch-codex-memory.md.

Preconditions

  • Prefer .memsearch/memory/ if it exists.
  • If there is no .memsearch/memory/, fall back to $agent-memory-ledger and rg across .codex-memory/.
  • If memsearch is not installed, stop and ask the user whether they want to install it. Do not run curl | sh installers.

Workflow

  1. Confirm the memory store location:

    • PROJECT_ROOT/.memsearch/memory/ (preferred)
    • otherwise .codex-memory/ + rg fallback
  2. Derive a stable project collection, then index memory when needed:

PROJECT_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
COLLECTION="$(plugins/codex-supercharge/scripts/derive_memsearch_collection.sh "$PROJECT_ROOT")"
memsearch index "$PROJECT_ROOT/.memsearch/memory" --collection "$COLLECTION"

If indexing fails because the ONNX embedding model needs a first-run download, either:

  • temporarily allow network access, or
  • switch to a provider you already have configured (high-trust only).
  1. Search for relevant chunks and expand the best hit:
memsearch search "<query>" --top-k 5 --json-output --collection "$COLLECTION"
memsearch expand <chunk_hash> --collection "$COLLECTION"
  1. If expand fails, read the source file directly using the source and line-range fields from JSON output:
sed -n '<start_line>,<end_line>p' "<source>"
  1. Return a curated summary:
    • focus on decisions, constraints, prior fixes, and “why”
    • cite the memory file/date (and heading if present)
    • if nothing relevant: say so, and propose a tighter query

Skip when

  • The task is purely about the current repo state (prefer Read/Grep/tests).
  • The user asked to ignore memory or avoid semantic tools.
  • There is no .memsearch/memory/ and the user does not want to create one.

Validation

  • If memsearch exists: memsearch --version and a dry run search against a small collection.
  • Otherwise: ensure the fallback path uses $agent-memory-ledger + rg.
  • Validate the helper with plugins/codex-supercharge/scripts/derive_memsearch_collection.sh "$PWD".