Recall

Load context from past sessions. Three modes: temporal (what did I work on yesterday/last week), topic (semantic search across sessions and notes), and graph...

Security Scan

The skill's purpose (loading past-session context) matches most of its instructions, but there are concrete inconsistencies and unclear behaviors — notably a referenced graph-generation script that doesn't exist in the package and vague filesystem paths — that make the bundle internally inconsistent and worth caution.

Purpose & Capability

The name/description (recall past sessions) aligns with instructions to read memory/*.md, memory/chat-log-*.jsonl, SESSION-STATE.md and synthesize a timeline. Reading those files is appropriate for the stated purpose. However, the SKILL.md references generating an HTML graph via python3 ${CLAUDE_SKILL_DIR}/scripts/generate-graph.py even though this is an instruction-only skill with no code files — that is an incoherence (the graph functionality as documented cannot run from this package as provided).

Instruction Scope

Runtime steps explicitly instruct the agent to run shell commands (ls, cat, grep) over workspace paths and to read 'any notes in the workspace' and SESSION-STATE.md. That general file-reading scope is expected for a recall skill, but it also means the agent will read arbitrary workspace files (which may contain secrets). The SKILL.md uses ambiguous path patterns (ls ~/[workspace]/memory/...) and an environment-like variable ${CLAUDE_SKILL_DIR} — these are underspecified and could lead to the agent searching unexpected locations. The graph step points to a script path that does not exist in the package, so the instruction set assumes external code that isn't provided.

Install Mechanism

There is no install spec and no code files, so nothing is written to disk by the skill itself. This is low install risk. The README mentions an optional third-party tool (@qmd/cli) that a user may install; installing that would introduce external code but is optional and documented.

Credentials

The skill declares no required environment variables, binaries, or credentials, which is proportional to its stated task. One minor inconsistency: instructions reference ${CLAUDE_SKILL_DIR} (an environment-like variable) but it is not declared or required — the agent may assume it exists in runtime, which is not guaranteed.

Persistence & Privilege

The skill does not request permanent presence (always:false) and is user-invocable. It does instruct the agent to read and write files in the user's workspace (e.g., outputting memory/recall-graph-YYYY-MM-DD.html), which is normal for this kind of skill but worth noting because it modifies workspace contents.

What to consider before installing

This skill appears to do what it claims (scan your workspace memory files and synthesize a timeline), but there are a few red flags you should consider before installing or enabling it: - Missing graph script: The SKILL.md calls python3 ${CLAUDE_SKILL_DIR}/scripts/generate-graph.py to build the visualization, but the package contains no scripts. If you need the graph feature you must supply that script or confirm where the agent should get it. - Broad file access: The skill's runtime instructions tell the agent to read memory/, SESSION-STATE.md, and 'any notes in the workspace' with shell commands (cat/grep). That will give the skill access to any files in those paths, which may include secrets or sensitive data. Consider running the skill only in a restricted/isolated workspace or reviewing files you keep under memory/ first. - Ambiguous paths and variables: The use of ~/[workspace], ${CLAUDE_SKILL_DIR}, and other underspecified locations could cause the agent to search unexpected directories. Confirm how your agent resolves those variables or edit the SKILL.md to use explicit, safe paths. - Optional dependencies: The doc suggests installing @qmd/cli for semantic search — installing global npm packages runs third-party code. Only install such tools from trusted sources. Recommendations: 1) Ask the skill author (or the publisher) for the missing graph script or remove/disable graph-related instructions. 2) Test the skill in a disposable or sanitized workspace (no secrets) first to observe exactly what files it reads/writes. 3) If you do not want automatic scanning of your entire workspace, avoid giving the agent autonomous invocation or restrict the skill's use to manual, user-invoked runs. 4) Consider adding an explicit allowlist of memory/ file paths in the SKILL.md before enabling it in a production workspace. Given the clear mismatch (documented functionality that depends on non-existent code and underspecified path handling), treat this package with caution until those inconsistencies are resolved.

Your agent forgets between sessions. Recall fixes that.

Three modes: temporal (date-based), topic (semantic search), graph (visual map). Every recall ends with One Thing — a concrete, specific next action synthesized from what you find.

---

Modes

1. Temporal — What were you working on?

/recall yesterday
/recall last week
/recall 2026-03-28
/recall this week
/recall last 3 days

Scans memory/YYYY-MM-DD.md and memory/chat-log-YYYY-MM-DD.jsonl files chronologically. Shows a timeline of sessions with topics, decisions made, and tasks in progress.

What it shows:

• Sessions list (time, topic, key decisions)
• Tasks that were in progress → which are done, which are stuck
• Decisions made and why
• Open loops that weren't closed

2. Topic — What do you know about X?

/recall authentication
/recall meeting with Bayram
/recall MoltNet architecture
/recall voice cloning

Searches across:

• memory/ daily files
• MEMORY.md long-term memory
• SESSION-STATE.md
• Any notes in the workspace

Uses keyword + semantic matching. Returns relevant excerpts with dates.

3. Graph — Visualize your work streams

/recall graph yesterday
/recall graph last week
/recall graph last 3 days

Generates an interactive HTML visualization:

• Sessions as nodes, colored by day
• Files touched as connected nodes
• Clusters reveal related work streams
• Shared files show cross-session dependencies

Output: memory/recall-graph-YYYY-MM-DD.html — opens in browser.

---

The One Thing

Every recall ends with synthesis:

Based on what has momentum, what's blocked, and what's closest to done — here's the single highest-leverage action right now:

[Specific, actionable, not generic]

Not "work on your project." More like "The auth flow has been blocked for 3 days — unblock it by making the decision about OAuth vs API keys you've been deferring."

---

How It Works

Step-by-step execution

Temporal query:

# 1. Find relevant memory files
ls ~/[workspace]/memory/YYYY-MM-DD.md  # today and requested range

# 2. Read them chronologically
cat memory/2026-03-28.md memory/2026-03-29.md ...

# 3. Also check SESSION-STATE.md for current hot context
cat memory/SESSION-STATE.md

# 4. Build timeline
# 5. Synthesize One Thing

Topic query:

# 1. Search memory files for topic
grep -r "TOPIC" memory/ --include="*.md" -l

# 2. If QMD is available (faster, semantic):
qmd search "TOPIC" -n 10

# 3. Read matched sections
# 4. Synthesize what's known + One Thing

Graph query:

# 1. Parse memory files for file references and decisions
# 2. Build adjacency data
# 3. Generate HTML with D3.js visualization
python3 ${CLAUDE_SKILL_DIR}/scripts/generate-graph.py --days 7 --output memory/recall-graph.html
# 4. Open in browser or share link

---

Configuration

No configuration required. Works with any workspace that has memory/ files.

Optional — better search with QMD:

# Install QMD for semantic search
npm install -g @qmd/cli
qmd index memory/ --collection sessions

Optional — richer graph with file tracking: Add to your AGENTS.md:

When touching important files, log them to memory/YYYY-MM-DD.md:
  - Touched: path/to/file.md

---

Examples

"What was I doing last week?"

📅 Last week (Mar 24–30):

Mon Mar 24: Voice cloning setup — installed Chatterbox, tested on DGX
Tue Mar 25: LanguageMirror backend — /clone-voice endpoint working
Wed Mar 26: Blocked on Russian TTS (CosyVoice failed), pivoted to English MVP
Thu Mar 27: Voxtral announcement — saved notes, compared to ElevenLabs
Fri Mar 28: Content pipeline — built preflight-check.sh, integrated with Moltbook

🔄 In progress: LanguageMirror end-to-end test (Pron)
✅ Done: Chatterbox, preflight-check
❌ Blocked: Russian TTS, Google Auth

⚡ One Thing: The LanguageMirror test has been queued for Pron for 3 days — check if it's done or unblock it.

"What do I know about the MoltNet architecture?"

📚 Found 8 references across 12 sessions (Mar 15 – Mar 31):

[2026-03-15] Decided: Node 1 Strasbourg as primary, Proxmox VE
[2026-03-18] Router deployed at /opt/moltnet/router/router.js
[2026-03-22] MoltWallet vs MoltPay distinction documented
[2026-03-28] Node 2 Singapore planned (Ryzen 9950X €299)

Key decision: Multi-tenant VPS marketplace, Telegram-first

⚡ One Thing: Node 2 Singapore has been "planned" for 2 weeks with no action — decide: this month or defer to Q3?

---

Upgrade Path

Setup	What you get
Just memory/ files	Temporal recall, basic topic search
+ QMD installed	Semantic topic search, faster
+ file tracking in AGENTS.md	Richer graph with file relationships
+ SESSION-STATE.md protocol	Real-time hot context always available

Start simple. Add as needed.

---

See Also

• evolutionary-model — why memory persistence matters
• onboarding — set up your memory foundation first
• ArtemXTech's original recall skill for Claude Code + Obsidian workflows