/arrecall
AgentRecall on-demand recall — surface past fixes, decisions, and patterns mid-session without leaving your flow.
$ npx -y skills add Goldentrii/AgentRecall-X --agent claude-codeHow it fires
How this command gets triggered: by you, by Claude, or both.
- Fires itselfClaude auto-loads it when your prompt matches the work.
- You can call itInvoke it directly when you want it.
- Slash command
/arrecall
Context preview
What this command does when you run it.
AgentRecall on-demand recall — surface past fixes, decisions, and patterns mid-session without leaving your flow.
Command definition
arrecall.mddescription: "AgentRecall on-demand recall — surface past fixes, decisions, and patterns mid-session without leaving your flow."
/arrecall — Contextual Mid-Session Recall
On-demand search against AgentRecall's memory store. Use anytime during a session — not just at start or end.
Two Modes
Mode A: Explicit Query (`/arrecall <query>`)
Direct recall. Run immediately:
recall({ project: "<current-project>", query: "<query>" })CLI equivalent if MCP tools aren't available: `ar recall "<query>"`.
Show results. Done. Use when you know exactly what to ask for.
Mode B: Contextual Scan (`/arrecall` with no args)
The useful mode. Extract context from the current conversation and run targeted multi-query recall automatically.
1. **Read the last 10 messages** of the current conversation 2. **Extract 3-5 keywords** from topics being discussed (technical components, decisions, domain terms) 3. **Run parallel recall queries** — one per most important keyword, plus one against the global/catchall project for cross-project awareness:
recall({ project: "<slug>", query: "keyword1" })
recall({ project: "<slug>", query: "keyword2" })
recall({ project: "<catchall-project>", query: "keyword1" })4. **Merge and deduplicate** results (same slug = same entry, skip duplicates) 5. **Show the compact hit list:**
──────────────────────────────────────
ArRecall — contextual hits
──────────────────────────────────────
[palace/decisions] We chose X over Y because... (2026-04-12)
[awareness] API returns null when session expires (6× confirmed)
[correction] Never push without permission [P0]
──────────────────────────────────────
3 results for: proxy auth, 407 error, headers
6. **If 0 results:** say "No hits for these topics — memory may not have captured this yet." and continue.
When to Invoke
Run `/arrecall` at these moments — don't wait to be asked:
- **Before any technical implementation** — query the component or domain you're about to touch
- **Before making an architectural decision** — query the decision topic
- **When something seems familiar** but details are fuzzy — trust the instinct, run the query
- **When the user says "last time" or "before"** — they're signaling that memory exists; go find it
- **When you hit a weird error** — check if it's been seen before
Important Rules
- **Never skip recall when the user references "before" or "last time"** — that's a direct signal relevant memory exists
- **Mode A is for moments of certainty; Mode B is for moments of uncertainty** — when you don't know what to ask, let the conversation guide the keywords
- **3 parallel calls is the default** in Mode B — not 1, not 5. Enough to triangulate without burning tokens
- **0 results is not a failure** — it means the topic hasn't been captured yet. Say so and continue
- **Cross-project query is always included** in Mode B — awareness insights apply across projects
---
Family: `/arstart` · `/arsave` · `/arrecall` · `/arreflect` — the four memory verbs (open · save · search · consolidate).
Read more
description: "AgentRecall on-demand recall — surface past fixes, decisions, and patterns mid-session without leaving your flow."
/arrecall — Contextual Mid-Session Recall
On-demand search against AgentRecall's memory store. Use anytime during a session — not just at start or end.
Two Modes
Mode A: Explicit Query (`/arrecall <query>`)
Direct recall. Run immediately:
recall({ project: "<current-project>", query: "<query>" })CLI equivalent if MCP tools aren't available: `ar recall "<query>"`.
Show results. Done. Use when you know exactly what to ask for.
Mode B: Contextual Scan (`/arrecall` with no args)
The useful mode. Extract context from the current conversation and run targeted multi-query recall automatically.
1. **Read the last 10 messages** of the current conversation 2. **Extract 3-5 keywords** from topics being discussed (technical components, decisions, domain terms) 3. **Run parallel recall queries** — one per most important keyword, plus one against the global/catchall project for cross-project awareness:
recall({ project: "<slug>", query: "keyword1" })
recall({ project: "<slug>", query: "keyword2" })
recall({ project: "<catchall-project>", query: "keyword1" })4. **Merge and deduplicate** results (same slug = same entry, skip duplicates) 5. **Show the compact hit list:**
────────────────────────────────────── ArRecall — contextual hits ────────────────────────────────────── [palace/decisions] We chose X over Y because... (2026-04-12) [awareness] API returns null when session expires (6× confirmed) [correction] Never push without permission [P0] ────────────────────────────────────── 3 results for: proxy auth, 407 error, headers
6. **If 0 results:** say "No hits for these topics — memory may not have captured this yet." and continue.
When to Invoke
Run `/arrecall` at these moments — don't wait to be asked:
- **Before any technical implementation** — query the component or domain you're about to touch
- **Before making an architectural decision** — query the decision topic
- **When something seems familiar** but details are fuzzy — trust the instinct, run the query
- **When the user says "last time" or "before"** — they're signaling that memory exists; go find it
- **When you hit a weird error** — check if it's been seen before
Important Rules
- **Never skip recall when the user references "before" or "last time"** — that's a direct signal relevant memory exists
- **Mode A is for moments of certainty; Mode B is for moments of uncertainty** — when you don't know what to ask, let the conversation guide the keywords
- **3 parallel calls is the default** in Mode B — not 1, not 5. Enough to triangulate without burning tokens
- **0 results is not a failure** — it means the topic hasn't been captured yet. Say so and continue
- **Cross-project query is always included** in Mode B — awareness insights apply across projects
---
Family: `/arstart` · `/arsave` · `/arrecall` · `/arreflect` — the four memory verbs (open · save · search · consolidate).
Correction-first persistent memory for AI agents. MCP server + SDK + CLI. Compounds across sessions.
Repo: Goldentrii/AgentRecall-X
Other commands on agentrecall-x.
- /arreflect
AgentRecall consolidation & reflection — periodic triage of recurring corrections; proposes rule changes, never applies them without the owner.
Open command - /arsave
AgentRecall full save — journal + palace + awareness + insights in one shot.
Open command - /arstart
AgentRecall session opener — status board (no args), project loader (<slug>), or cold-start bootstrap. Use first every session.
Open command

