Memory Tips

Build Memory Deliberately

The agent auto-extracts simple preferences from your messages, but the best memories are explicit:

memory add fact "Our main customer is Acme Corp, renewal date Q3 2026, ARR $120k"
memory add fact "The engineering team uses GitHub for code review, Jira for project tracking"
memory add preference "Always format reports as bullet points, not prose"

Use Personas for Context Switching

Create named personas to keep work and personal contexts separate:

memory persona_set work "You are my work assistant. Focus on enterprise software sales."
memory persona_set personal "You are my personal assistant. I live in London, prefer brief responses."

Activate a persona for a session by including it in your first message, or set agentName via /template use.

Share Memories with Your Team

Memories default to private. To share a key insight with your whole team:

PATCH /v1/memories/:id
{ "visibility": "team" }

Or in the web dashboard: Memory → click the memory → Share with team.

Semantic Search Works Best With Nouns

The cosine similarity search finds memories most similar to your query. Queries with specific nouns work better than vague ones:

Good:  "Acme Corp renewal"
Good:  "engineering team tools"
Weak:  "work stuff"

Backfill Old Memories

If you added memories before configuring an embedder, backfill them:

POST /v1/jobs
{ "kind": "embed-memories", "payload": { "limit": 500 } }

Check the job status at GET /v1/jobs?status=completed.

Memory Lifecycle

• Memories are permanent until you delete them
• The agent never deletes memories automatically
• Use memory list kind=preference regularly to prune stale preferences
• The learning loop adds skill-note kind memories — these are automatically generated and can get verbose; prune them monthly