Skip to content
Agent Memory
Skill

/recall

Memory recall guide. Structured retrieval from hipocampus memory — ROOT.md triage, manifest-based LLM selection, qmd search fallback.

From plugin
hipocampus
1595 skills3 hooks
Install
$ npx -y skills add kevin-hs-sohn/hipocampus --skill recall --agent claude-code

How it fires

How this skill gets triggered: by you, by Claude, or both.

  • Fires itselfAuto-invocation. Claude auto-loads it when your prompt matches the work.Auto-invocation is when the right skill fires by itself at the right moment, driven by a FLOW.md router and a hook, instead of you invoking it by name. It is the difference between a skill being installed and a skill actually getting used.Read the full definition →
  • You can call itInvoke it directly when you want it.
  • Slash command/recall

Context preview

The summary Claude sees to decide when to auto-load this skill.

Memory recall guide. Structured retrieval from hipocampus memory — ROOT.md triage, manifest-based LLM selection, qmd search fallback.

SKILL.md

recall.SKILL.md
name: hipocampus-recall
description: "Memory recall guide. Structured retrieval from hipocampus memory — ROOT.md triage, manifest-based LLM selection, qmd search fallback."

Memory Recall Protocol

Use this when the user's question may relate to past memory. Three-step fallback: ROOT.md O(1) lookup → manifest LLM selection → qmd search.

Step 1: ROOT.md Triage (O(1) — always try first)

Check ROOT.md Topics Index for the query topic.

  • **Direct match found** → read the referenced file (knowledge/, daily log date, etc.). Done.
  • **Partial match / related topic found** → read referenced file, check if it answers the query. Done if yes.
  • **No match at all** → proceed to Step 2.

**Decision rule:** If Topics Index contains a keyword within 1 semantic hop of the query, it's a match. "배포" matches "deployment". "CI/CD" matches "github-actions".

Step 2: Manifest-Based LLM Selection (when ROOT.md is insufficient)

Use this ONLY when ROOT.md Topics Index has no relevant match but you suspect memory may exist (e.g., the user references something that sounds familiar, or the topic is cross-domain).

1. **Build manifest** from compaction node frontmatter (NOT full content):

  • Read `memory/weekly/*.md` frontmatter only (type, period, topics)
  • Read `memory/monthly/*.md` frontmatter only (type, period, topics)
  • Read `knowledge/*.md` first 3 lines only
  • Skip `memory/daily/` (already rolled up into weekly)

2. **Self-evaluate:** Given the manifest and the user's query, select up to 5 most relevant files.

3. **Load selected files** in full and extract the answer.

**Token budget:** Manifest should be <500 tokens. If too large, use monthly nodes only.

Step 3: qmd Search (fallback)

If Step 1-2 don't find the answer and qmd is installed:

qmd query "keyword1 keyword2"      # hybrid (BM25 + vector)
qmd search "keyword1 keyword2"     # BM25 only
qmd vsearch "semantic query"       # vector only

Use 2-4 specific keywords. Try variations if first query misses.

Freshness Warnings

When recalling memory, check the source age:

  • `project` type + >30 days old: append warning — "이 정보는 {N}일 전 기록입니다. 현재 상태를 확인하세요."
  • `reference` type + `[?]` marker: append warning — "이 참조는 검증되지 않았습니다. 접근 가능 여부를 확인하세요."
  • `user`/`feedback` type: no age warning (these are durable).

When NOT to Use Recall

  • If the answer is clearly in ROOT.md Active Context → just use it directly
  • If the question is about code/files → read the code, don't search memory
  • If the question is about git history → use git log, not memory

Key Principles

  • ROOT.md O(1) triage is the primary path — Step 2/3 are fallbacks
  • Never load full content in Step 2 manifest phase — frontmatter only
  • Manifest LLM selection leverages semantic understanding (solves keyword mismatch)
  • Minimize token usage: most recalls should resolve at Step 1
Read more
Ships withhipocampus

Drop-in proactive memory harness for AI agents. Zero infrastructure — just files. One command to set up. Works immediately with Claude Code, OpenCode, and OpenClaw.

Get the whole plugin
Stats
191
Stars
9
Forks
Maintained
Maintenance
JavaScript
Language
MIT
License
2mo ago
Last commit
5mo ago
Created

Repo: kevin-hs-sohn/hipocampus