Build Memory

Build Memory — Turn Local Context Into Agent Memory

Your agent's memory is notes under /Memory. /Memory/Self/ holds the identity files that shape how the agent thinks and what it knows:

• USER.md — who the owner is (role, stack, current work, expertise, goals)
• COO.md — the agent's voice and operating style
• POLICY.md — boundaries (what to share, what never to share)
• MEMORY.md — long-term facts, projects, preferences, learned patterns

Building memory = reading the user's real context, synthesizing it, and writing these files. This is more powerful than a generic importer because you (the agent) can read the actual repo, git history, and docs and reason about what matters.

Requires being signed in — see the onboarding skill. Resolve the token first with the pack's scripts/aicoo-auth.sh (auto-refreshes): export AICOO_API_KEY="$(bash /path/to/aicoo-skills/scripts/aicoo-auth.sh)".

1. Ask where to build from

Don't guess the source — ask, then use what they pick:

"Where should I build your memory from? I can use: • this folder / repo (I'll read the code, docs, and git history) • a different local folder — tell me the path • Notion or Google Docs — I'll point you to the importer • just tell me about yourself — we skip files"

• Local folder / repo / git → you (the agent) read it directly (step 2).
• Notion / Google Docs → these connect through the web importer, which the skill can't call with an API token. Point the user to https://www.aicoo.io/import-workspace (Connect Notion / Google Drive → select docs → it extracts memory + generates USER.md/COO.md). Then come back and continue from step 4 to verify.
• Just tell me → ask 3–4 questions (role, what they build, stack, how they want the agent to behave) and synthesize from that.

2. Scan the chosen local source

Pull the strongest available signals; adapt to what exists (use the folder the user named, or the current one):

cat README.md package.json pyproject.toml Cargo.toml 2>/dev/null | head -200
git log --oneline -30 2>/dev/null
git shortlog -sn 2>/dev/null | head
ls docs/ notes/ 2>/dev/null
cat CLAUDE.md AGENTS.md .claude/*.md 2>/dev/null | head -200

Read domain/architecture/decision docs. If signals are thin, ask the user 2–3 targeted questions (role, what they're building, how they want their agent to behave) rather than writing empty files.

3. Synthesize

Draft the four identity files from what you found. Be concrete and specific — real project names, real stack, real facts. MEMORY.md should be a running list of durable facts/projects/preferences, not prose.

Structure MEMORY.md like:

# Long-term Memory
## Key facts
- <fact>
## Projects
- <project>: <one line>
## Preferences
- <how they like to work / how the agent should behave>

4. Write it to Aicoo

accumulate with Memory/... paths routes into the memory space. It create-or-updates (with versioning), so this skill is safe to re-run to refresh memory later:

curl -s -X POST "https://www.aicoo.io/api/v1/accumulate" \
  -H "Authorization: Bearer $AICOO_API_KEY" -H "Content-Type: application/json" \
  -d '{
    "files": [
      {"path": "Memory/Self/USER.md",   "content": "..."},
      {"path": "Memory/Self/COO.md",    "content": "..."},
      {"path": "Memory/Self/POLICY.md", "content": "..."},
      {"path": "Memory/Self/MEMORY.md", "content": "..."}
    ]
  }' | jq '{created, updated, skipped, workspace}'

For a team knowledge base, write topical notes into their own folder instead of Memory/Self, e.g. {"path": "Knowledge Base/architecture.md", ...} — then the onboarding / share-agent skills can share that folder as a unit.

5. Verify (the aha)

curl -s -X POST "https://www.aicoo.io/api/v1/os/memory/search" \
  -H "Authorization: Bearer $AICOO_API_KEY" -H "Content-Type: application/json" \
  -d '{"query": "what am I working on", "maxResults": 5}' | jq '.result.results[] | {title, snippet, path}'

Show the user that their agent can now recall what you just wrote. To make it tangible, create a share link (share-agent skill) and talk to their own agent as a guest — it answers from this memory.

Refresh later

Re-running this skill overwrites the identity files with a fresh synthesis (auto-versioned, so history is kept). For incremental updates that append rather than replace, use the context-sync / autonomous-sync skills.

Notes

• accumulate tags these as workspace context; they live at the correct /Memory/Self path and are readable via os/memory/search and os/notes.
• Snapshot before a big rewrite if you want a restore point: POST /api/v1/os/snapshots/{noteId} (see the snapshots skill).
• Never write secrets/keys into memory. Put sharing boundaries in POLICY.md.