Appearance
Memory
Memory is Pencel's knowledge system — it lets agents remember what they've learned across runs and conversations.
Where memory lives
You review and curate memory under Settings → Memory. Agents draw on it automatically as part of Knowledge when they run.
What It Does
Without memory, every agent run starts from scratch. Memory changes that. After each run, Pencel automatically summarizes key learnings and stores them. When an agent starts a new task, relevant memories are retrieved and included in the agent's context — so it builds on what it already knows.
Over time, individual memories are distilled into broader patterns and preferences. Your agents get smarter the more they work.
Memory Types
Pencel organizes knowledge into four memory types, each serving a different purpose.
| Type | What It Stores | Example |
|---|---|---|
| Episodic | Key takeaways from individual runs | "Q3 revenue was $2.4M, up 12% from Q2" |
| Long-term | Distilled patterns and enduring facts across runs | "The finance team prefers bullet-point summaries" |
| Preference | User and organizational preferences | "Always CC the VP of Sales on revenue reports" |
| Pattern | Recurring behaviors or trends | "Support tickets spike every Monday morning" |
TIP
You do not need to create memories manually. Pencel generates episodic memories automatically after each run. You can add long-term topic notes and similar knowledge from Settings → Memory to give agents a head start. Preference memories are created when you confirm corrections from chat; pattern memories are extracted automatically from tool use (e.g. MCP).
How Memory Builds Over Time
Memory in Pencel follows a natural learning cycle:
- Capture — After each run, key learnings are automatically summarized into episodic memories.
- Retrieve — When an agent starts a new task, Pencel finds memories that are semantically similar to the current task and injects them into the agent's context.
- Distill — Over time, related episodic memories are consolidated into long-term memories that capture broader patterns.
- Apply — Agents use these long-term memories to make better decisions without needing to re-learn the same lessons.
INFO
Memory retrieval uses semantic similarity, not keyword matching. If an agent is working on a "quarterly earnings summary," it will find memories about "Q3 revenue" even though the words are different.
How Distillation Works
Distillation is what turns a pile of episodic memories into durable knowledge your agents can actually use. It happens automatically in two passes.
- Incremental consolidation — after each run completes, the agent reviews what just happened and pulls out anything worth keeping: a preference confirmed, a pattern observed, or durable facts. These become new long-term, preference, or pattern memories on the spot, without you doing anything.
- Weekly sweep — once a week, Pencel does a full pass over recent episodic memories. It deduplicates near-identical entries, merges related observations, and promotes patterns that have shown up consistently into higher-confidence long-term memories. One-off observations stay episodic and eventually fade.
- What gets kept — durable facts, repeated patterns, and explicit preferences. Throwaway details from a single run (a one-time API rate limit, a transient error) get filed as episodic and don't survive the sweep.
- Decay and cleanup — unused, low-value memories decay over time and are garbage-collected, so the pool stays lean. A first-run grace period avoids pruning a fresh workspace before it has had a chance to learn.
TIP
You can speed up the learning cycle by adding long-term knowledge directly in Settings → Memory — useful for facts you already know (team conventions, naming rules, who to CC) that you do not want to wait for the agent to learn.
See Using Memory for hands-on instructions on adding, editing, and reviewing memories.
Key Properties
| Property | Description |
|---|---|
| Content | The text of the memory |
| Type | One of the four memory types above |
| Source | Where the memory came from (run ID, session, or manual) |
| Confidence | Score from 0 to 1 indicating reliability |
| Usage Count | How many times this memory has been retrieved |
| Enabled | Whether the memory is active or disabled |
| Category | Optional grouping label |
Managing Memories
You can review, edit, and manage memories from Settings → Memory.
- Edit a memory to correct inaccurate information.
- Disable a memory to stop it from being included in agent context without deleting it.
- Delete outdated memories that are no longer relevant.
- Boost confidence on memories you have verified to be accurate.
WARNING
Memories with low confidence scores are still included when relevant. If you notice an agent using incorrect information, check Settings → Memory for outdated or wrong memories and disable them.
What's Next
- Context Files — provide static background knowledge alongside dynamic memory
- Runs — see which memories an agent used during a run
- Agents — each agent draws from the same shared memory pool