Appearance
Seed Your Agent's Memory
Add durable knowledge so your agent starts with the context it needs, then watch it apply that knowledge in conversation.
What You'll Use
| Feature | Purpose |
|---|---|
| Memory | Persistent knowledge that agents draw on automatically |
| Memory Types | Different categories for different kinds of knowledge |
| Chat | Test that the agent retrieves and uses your memories |
Step 1: Add a long-term note (formatting guidelines)
Tell the agent how you like summaries written.
- Go to Knowledge in the sidebar, open the Memory tab.
- Stay on the Knowledge sub-tab (not Journal).
- Click New Note (creates a new long-term memory item).
| Field | Value |
|---|---|
| Content | Leadership prefers bullet-point summaries over narrative paragraphs. Keep summaries under 10 bullet points. Lead with the most important item. Use bold for key metrics. |
Save when done.
Step 2: Add project facts (second long-term note)
Capture enduring facts about a project the same way.
- Click New Note again.
| Field | Value |
|---|---|
| Content | Project Phoenix — internal initiative launched January 2025, led by Sarah Chen (VP Engineering). Goal: migrate analytics from on-prem Hadoop to BigQuery. Status: Phase 2 (data migration). Budget: $1.2M. Target completion: Q3 2025. Key risk: legacy ETL pipelines with undocumented transformations. |
INFO
Memory retrieval is semantic, not keyword-based. The agent does not need the exact phrase "Project Phoenix" — asking about "the analytics migration" or "Sarah Chen's project" can still match.
Step 3: Test memory retrieval
- Open the Chat panel and select an agent.
- Ask:
"Write a status update on the analytics migration project."
The agent should use your long-term notes: bullet-style guidance plus Phoenix details (Sarah Chen, BigQuery, Phase 2, budget).
- Now ask something unrelated:
"What's a good framework for running a retrospective?"
The agent should answer without leaning on Phoenix-specific facts — retrieval is relevance-based.
Memory types at a glance
| Type | Best for | How it's created |
|---|---|---|
| Episodic | What happened during a specific run or conversation | Automatic (runs / chat) |
| Long-term | Enduring knowledge (policies, project facts, how you like outputs) | Memory view (e.g. New Note, topic docs) or distillation |
| Preference | Explicit "how we work" preferences tied to corrections | Automatic (chat correction flow) |
| Pattern | Recurring technical observations (e.g. MCP / API usage) | Automatic |
You typically seed long-term knowledge manually from the Memory view; episodic, preference, and pattern memories accumulate as you use Pencel.
TIP
After important runs, check the Journal sub-tab under Memory — episodic captures often appear there. Use Extract Learnings on the Memory tab to roll journal content into topic-based long-term knowledge.
Step 4: Maintain your memories
Memories can go stale. Periodically review them:
- Edit when facts change (e.g. project status).
- Disable items that are temporarily irrelevant.
- Delete items that are wrong or obsolete.
Low-confidence memories are still used but ranked lower in retrieval. Adjust confidence as you verify accuracy.
What to try next
- Memory — How semantic retrieval works, confidence scoring
- Using Memory — Reviewing, editing, and curating memories
- Teach Your Agent About Your Company — Use context files for static reference material