/arsave

description: "AgentRecall full save — journal + palace + awareness + insights in one shot."

/arsave — AgentRecall Full Save

One command to save everything. No long prompts needed.

When to Use

**Default: USE IT.** Most projects are long-term. Memory compounds — insights saved today prevent repeated mistakes and rebuild costs across future sessions.

**Skip /arsave only when** the session was truly throwaway:

• Pure Q&A with no decisions made
• Trivial one-off task that won't be revisited
• Nothing non-obvious happened worth recalling

What This Does

Runs the complete AgentRecall end-of-session flow:

1. **Gather** — review what happened this session 2. **Save** — one `session_end` call writes journal + awareness + consolidation 3. **Verify** — check that key content was promoted 4. **Git** — push to GitHub if user has configured it

Process

Step 1: Gather session context

**Start with machine-captured facts, not memory.** At long context windows your memory of early decisions is compressed and unreliable. Ground truth comes first:

1. **Read today's capture log** — `~/.agent-recall/projects/<slug>/journal/YYYY-MM-DD-log.md` (if it exists). This file contains incremental Q&A captures logged during the session. Pull out the key facts from it.

2. **Check git diff** — if in a git repo, run `git diff --stat HEAD` or `git log --oneline -5` to see what files actually changed.

3. **Supplement with memory** — now recall what happened that isn't in the log: decisions made in conversation, things we discussed but didn't act on, blockers identified, next steps.

Combine all three into a 2-3 sentence summary. The log anchors you; memory fills the gaps.

Step 1b: Capture intention on first save

Check whether this project already has an intention recorded:

grep -l "Intention:" ~/.agent-recall/projects/<slug>/palace/identity.md 2>/dev/null

**If NOT found** (this is the first save for this project, or intention was never captured):

Look at the earliest user messages in this conversation — where the user explained what they're trying to do, why they're starting this project, what problem they're solving, or what their goal is. Extract one clear sentence that captures the core WHY.

Write it to the identity file:

mkdir -p ~/.agent-recall/projects/<slug>/palace
# prepend to identity.md (or create it)
echo "**Intention:** <extracted intention sentence>" | cat - ~/.agent-recall/projects/<slug>/palace/identity.md 2>/dev/null > /tmp/identity-tmp.md && mv /tmp/identity-tmp.md ~/.agent-recall/projects/<slug>/palace/identity.md

Rules for extraction:

• One sentence, max ~20 words
• Capture the WHY and WHAT, not the HOW (e.g. "Build a pixel-perfect replica of novada.com as a deployable Next.js site")
• If no clear intention is detectable from the conversation (e.g. pure Q&A, no project context), skip — don't fabricate

**If already found**: skip this step entirely.

Step 2: Record corrections

If the human corrected your understanding during this session — "no not that", "I meant X not Y", "wrong priority" — record each significant correction:

check({
  goal: "<what you originally understood>",
  confidence: "high",
  human_correction: "<what the human actually wanted>",
  delta: "<the gap — e.g. 'assumed technical priority, human meant business priority'>"
})

This feeds the predictive warning system. Future agents on this project will get `watch_for` warnings like: "You tend to misinterpret X — corrected N times."

If no corrections happened this session, skip this step.

Step 3: Save everything in one call

Call `session_end` with:

session_end({
  summary: "<2-3 sentence session summary>",
  insights: [
    {
      title: "<one-line insight>",
      evidence: "<what happened that confirmed this>",
      applies_when: ["keyword1", "keyword2"],
      severity: "critical" | "important" | "minor"
    }
    // 1-3 insights max
  ],
  trajectory: "<where is the work heading — one line>"
})

This single call:

• Writes the daily journal entry
• Updates awareness with new insights (merge or add)
• Consolidates decisions/goals/blockers into palace rooms
• Archives demoted insights (not deleted — moved to awareness-archive.json)

Step 4: Verify promotion

After `session_end`, verify that content actually made it to the right places:

1. Call `recall(query="<key decision from today>")` — confirm it appears in palace results 2. Check the session_end response: `insights_processed` should match what you sent

If gaps found, use `remember` to manually save the missing content:

remember({
  content: "<the missing decision/insight>",
  context: "architecture decision"  // hints the router
})

Step 5: Output the save card

Render the following card. Replace all `<placeholders>` with real values from the session_end response and the actual project slug. Count the project's journal files to get the session number (`ls ~/.agent-recall/projects/<slug>/journal/*.md 2>/dev/null | wc -l`).

──────────────────────────────────────────────────────────────
  AgentRecall  ✓ Saved    <project-slug>   <YYYY-MM-DD>   #<N>
──────────────────────────────────────────────────────────────
  Intention     palace/identity.md
                └─ "<captured intention>"           [captured]   ← only if written this save

  Journal       ~/.agent-recall/projects/<slug>/journal/
                └─ <YYYY-MM-DD>.md                    [written]

  Awareness     ~/.agent-recall/awareness.md
                └─ <N> insights added  (<M> total)

  Palace        ~/.agent-recall/projects/<slug>/palace/
                ├─ rooms/<room1>.md                   [updated]
                ├─ rooms/<room2>.md                   [updated]
                └─ palace-index.json                 [reindexed]

  Insights      ~/.agent-recall/insights-index.json
                └─ cross-project index updated
──────────────────────────────────────────────────────────────

Omit