Pencel

Appearance

Using Memory ​

Memory is how your Pencel workspace builds institutional knowledge over time. Agents learn from completed runs, capture important patterns, and recall relevant context when working on future tasks. This guide explains how memory works and how you can curate it.

How memory works ​

Every time an agent completes a run or finishes a meaningful conversation, Pencel can automatically capture what it learned as a memory item. Over time, your workspace accumulates a library of experiences, preferences, patterns, and facts that agents draw on when performing tasks.

Memory is local to your workspace. It never leaves your machine. Agents retrieve relevant memories automatically when working on tasks — they search for memories related to the current context and weave that knowledge into their reasoning.

Memory types ​

Pencel organizes memories into four types, each serving a different purpose:

TypeWhat it capturesExample
EpisodicA summary of a specific run or conversation — what happened, what was decided, what worked"Monthly report run on Jan 15: pulled CRM data, found 12% drop in pipeline. Recommended increasing outbound calls."
Long-termEnduring facts or knowledge that remain true over time"Our fiscal year starts in April. Q1 = Apr-Jun, Q2 = Jul-Sep."
PreferenceUser or team preferences about how things should be done"Leadership prefers bullet-point summaries over narrative paragraphs."
PatternRecurring trends or behaviors the agent has observed"Support ticket volume spikes every Monday morning, usually related to weekend deployments."

INFO

Episodic memories are created automatically after runs. Long-term memories can be added from the Memory view or produced by distillation. Preference memories are created when you confirm corrections from chat. Pattern memories are extracted automatically from tool interactions (e.g. MCP).

Browsing memories ​

Open the Memory view ​

Navigate to Settings → Memory. You see a list of all memory items in your workspace.

Use the search bar to find memories by keyword. You can also filter by:

  • Type — Show only episodic, long-term, preference, or other specific types.
  • Date range — Find memories created within a specific time period.
  • Agent — See memories captured by a particular agent.

Read memory details ​

Click on any memory item to see its full content, including:

  • The memory text itself.
  • When it was created.
  • Which agent or run created it.
  • The confidence score (see below).
  • Whether it is active or disabled.

Understanding confidence scores ​

Each memory has a confidence score that indicates how reliable or relevant the information is. Scores range from low to high:

  • High confidence — The information comes from a verified source or has been confirmed multiple times. Agents prioritize high-confidence memories.
  • Medium confidence — The information is likely accurate but has not been independently verified. Agents use it but may note the uncertainty.
  • Low confidence — The information is tentative — perhaps inferred from a single data point or an ambiguous conversation. Agents treat it as a suggestion rather than a fact.

Confidence scores are set automatically based on how the memory was created. You can manually adjust them if you know the information is more or less reliable than the automatic score suggests.

Editing memories ​

You can edit any memory item to correct, update, or refine its content:

  1. Open the memory by clicking on it.
  2. Edit the text to fix inaccuracies or add context.
  3. Adjust the confidence score if needed.
  4. Save your changes.

TIP

Regularly reviewing and editing memories is one of the highest-value activities you can do in Pencel. Clean, accurate memories lead to better agent performance across every task.

Disabling and deleting memories ​

Sometimes a memory becomes outdated or irrelevant:

  • Disable — The memory remains in your workspace but agents no longer retrieve it. Use this for information that might be useful again in the future, like seasonal patterns.
  • Delete — The memory is permanently removed. Use this for incorrect information that should never influence agent behavior.

To disable or delete, open the memory item and use the status toggle or delete button.

How agents retrieve memories ​

When an agent starts a task, Pencel automatically searches for relevant memories using semantic matching. Here is what happens behind the scenes:

  1. Pencel looks at the current task context — the workflow instructions, the user's message, or the current step.
  2. It searches the memory library for items that are semantically similar to the current context.
  3. The top matches (up to five) are included in the agent's working context.
  4. The agent uses these memories alongside its instructions, guidelines, and any connected data.

This means agents do not retrieve every memory on every task. They only get the memories most relevant to what they are doing right now, keeping their focus sharp and their context efficient.

WARNING

Agents can only retrieve memories that are active. If you disable a memory, it will not appear in any agent's context regardless of relevance.

Automatic memory capture ​

After a workflow run completes, Pencel can automatically summarize the key takeaways and save them as episodic memories. This happens in the background — you do not need to do anything.

Automatic capture records:

  • What the agent did during the run.
  • Key decisions and their reasoning.
  • Results and outcomes.
  • Any issues or errors encountered.

Over time, this creates a rich history that agents can draw on. For example, an agent generating this month's report can recall what happened during last month's report run, including any adjustments that were made.

Tips for curating workspace knowledge ​

  • Review episodic memories after important runs. Check that the automatic summary captured the right takeaways. Edit if needed.
  • Add long-term memories manually for critical facts. Things like fiscal year dates, key contacts, naming conventions, and standard processes should be explicit long-term memories, not left to chance.
  • Clean up regularly. Every few weeks, browse your memories and disable or delete items that are outdated. Stale memories can mislead agents.
  • Use entity memories for key accounts and projects. Store essential details about important clients, systems, or initiatives so agents always have that context available.
  • Prefer specific over general. A memory that says "Client X prefers bi-weekly updates sent on Fridays by 3 PM" is far more useful than "Some clients want regular updates."
  • Watch for contradictions. If two memories conflict, agents may behave inconsistently. Resolve contradictions by editing or deleting the outdated item.

Next steps ​