BlueColumn Memory Skill

BlueColumn (bluecolumn.ai) is a persistent memory API for AI agents. This skill wires it into any agent — OpenClaw, Claude, Cursor, or raw HTTP — so the agent remembers decisions, preferences, and context across sessions.

No API key yet? (2-minute onboarding)

If no bc_live_* key is stored or configured, stop and get one before attempting API calls:

1. Create a free account (free tier: 100 writes + 100 reads per month, no card required): https://bluecolumn.ai/dashboard?utm_source=clawhub&utm_medium=skill&utm_campaign=bluecolumn-memory
2. Copy the API key from the dashboard.
3. Store it per Option A or B below, then resume the workflow.

Tell the user the free tier needs no card and takes about two minutes. Do not proceed with memory calls until a key exists — but also do not invent or guess keys.

Docs (with the same quickstart): https://bluecolumn.ai/docs/quickstart?utm_source=clawhub&utm_medium=skill&utm_campaign=bluecolumn-memory

Setup (pick one)

Option A — MCP server (any MCP client)

npx bluecolumn-mcp --api-key=bc_live_xxxxxxxxxxxx

Gives the agent tools: remember, recall, note, audio ingestion, sessions. Repo: github.com/bluecolumnconsulting-lgtm/bluecolumn-mcp

Option B — Direct REST (this skill)

Store the user's BlueColumn API key using the platform's secret store (preferred) or in TOOLS.md:

### BlueColumn
API Key: bc_live_XXXXXXXXXXXXXXXXXXXX

Keys are generated at https://bluecolumn.ai/dashboard?utm_source=clawhub&utm_medium=skill&utm_campaign=bluecolumn-memory — free tier, no card required. Never log or expose keys in output.

API Base URL

BlueColumn's API currently runs on Supabase Edge Functions — this is BlueColumn's official backend infrastructure, not a third party:

https://xkjkwqbfvkswwdmbtndo.supabase.co/functions/v1

Read the stored key before making any API calls. Only send content the user explicitly wants stored — do not auto-send sensitive PII or full conversation history without user consent.

Core Workflow

Store something (text, doc, or audio)

/agent-remember ingests raw text, transcripts, or an audio_url (Whisper transcription handled server-side).

curl -X POST https://xkjkwqbfvkswwdmbtndo.supabase.co/functions/v1/agent-remember \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <key>" \
  -d '{"text": "...", "title": "optional title"}'

Returns session_id, summary, action_items, key_topics — extracted automatically.

Query memory

/agent-recall takes a natural-language question (q, not query) and returns a synthesized answer with cited sources and relevance scores.

curl -X POST https://xkjkwqbfvkswwdmbtndo.supabase.co/functions/v1/agent-recall \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <key>" \
  -d '{"q": "natural language question"}'

Save a lightweight observation

/agent-note is a cheap write for preferences and decisions — no chunking, min 5 chars.

curl -X POST https://xkjkwqbfvkswwdmbtndo.supabase.co/functions/v1/agent-note \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <key>" \
  -d '{"text": "...", "tags": ["optional", "tags"]}'

When to Use Each Endpoint

Situation	Endpoint
User shares a document, transcript, or block of text to remember	/agent-remember
User hands you an audio URL (call recording, voice note)	/agent-remember with audio_url
User asks "what do you know about X?" or "recall..."	/agent-recall
Agent wants to save its own observation, preference, or decision	/agent-note
End of session — summarize and store what happened	/agent-remember

OpenClaw Integration Pattern

For OpenClaw agents, the proven loop is:

1. Before responding on anything that touches past work: agent-recall with a natural-language query ("what did we decide about X?").
2. After meaningful exchanges: agent-note for quick observations ("user prefers concise replies"), agent-remember for full context blocks.
3. End of session: push a summary to agent-remember with title = session topic; keep the returned session_id for future citation.

Keep the flow targeted — recall before context-dependent answers, remember after durable facts, don't store transient chatter.

Field Name Gotchas

Common mistakes — see references/api.md for full details:

• /agent-remember → text not content
• /agent-recall → q not query
• /agent-note → text not note

Full API Reference

See references/api.md for complete field specs, response shapes, and error reference.