Penfield Memory

Persistent memory that compounds. Your agent remembers conversations, learns preferences, connects ideas, and picks up exactly where it left off—across sessions, days, and channels.

Tools

Memory

Tool	Purpose	When to use
penfield_store	Save a memory	User shares preferences, you make a discovery, a decision is made, you learn something worth keeping
penfield_recall	Hybrid search (BM25 + vector + graph)	Need context before responding, resuming a topic, looking up prior decisions
penfield_search	Semantic search (higher vector weight)	Fuzzy concept search when you don't have exact terms
penfield_fetch	Get memory by ID	Following up on a specific memory from recall results
penfield_update_memory	Edit existing memory	Correcting, adding detail, changing importance or tags

Knowledge Graph

Tool	Purpose	When to use
penfield_connect	Link two memories	New info relates to existing knowledge, building understanding over time
penfield_explore	Traverse graph from a memory	Understanding how ideas connect, finding related context

Context & Analysis

Tool	Purpose	When to use
penfield_save_context	Checkpoint a session	Ending substantive work, preparing for handoff to another agent
penfield_restore_context	Resume from checkpoint	Picking up where you or another agent left off
penfield_list_contexts	List saved checkpoints	Finding previous sessions to resume
penfield_reflect	Analyze memory patterns	Session start orientation, finding themes, spotting gaps

Artifacts

Tool	Purpose	When to use
penfield_save_artifact	Store a file	Saving diagrams, notes, code, reference docs
penfield_retrieve_artifact	Get a file	Loading previously saved work
penfield_list_artifacts	List stored files	Browsing saved artifacts
penfield_delete_artifact	Remove a file	Cleaning up outdated artifacts

Writing Memories That Actually Work

Memory content quality determines whether Penfield is useful or useless. The difference is specificity and context.

Bad — vague, no context, unfindable later:

"User likes Python"

Good — specific, contextual, findable:

"[Preferences] User prefers Python over JavaScript for backend work.
Reason: frustrated by JS callback patterns and lack of type safety.
Values type hints and explicit error handling. Uses FastAPI for APIs."

What makes a memory findable:

1. Context prefix in brackets: [Preferences], [Project: API Redesign], [Investigation: Payment Bug], [Decision]
2. The "why" behind the "what" — rationale matters more than the fact itself
3. Specific details — names, numbers, dates, versions, not vague summaries
4. References to related memories — "This builds on [earlier finding about X]" or "Contradicts previous assumption that Y"

Memory Types

Use the correct type. The system uses these for filtering and analysis.

Type	Use for	Example
fact	Verified, durable information	"User's company runs Kubernetes on AWS EKS"
insight	Patterns or realizations	"Deployment failures correlate with Friday releases"
correction	Fixing prior understanding	"CORRECTION: The timeout isn't Redis — it's a hardcoded batch limit"
conversation	Session summaries, notable exchanges	"Discussed migration strategy. User leaning toward incremental approach"
reference	Source material, citations	"RFC 8628 defines Device Code Flow for OAuth on input-constrained devices"
task	Work items, action items	"TODO: Benchmark recall latency after index rebuild"
strategy	Approaches, methods, plans	"For user's codebase: always check types.ts first, it's the source of truth"
checkpoint	Milestone states	"Project at 80% — auth complete, UI remaining"
identity_core	Immutable identity facts	Set via personality config, rarely stored manually
personality_trait	Behavioral patterns	Set via personality config, rarely stored manually
relationship	Entity connections	"User works with Chad Schultz on cybersecurity content"

Importance Scores

Use the full range. Not everything is 0.5.

Score	Meaning	Example
0.9–1.0	Critical — never forget	Architecture decisions, hard-won corrections, core preferences
0.7–0.8	Important — reference often	Project context, key facts about user's work
0.5–0.6	Normal — useful context	General preferences, session summaries
0.3–0.4	Minor — background detail	Tangential facts, low-stakes observations
0.1–0.2	Trivial — probably don't store	If you're questioning whether to store it, don't

Connecting Memories

Connections are what make Penfield powerful. An isolated memory is just a note. A connected memory is understanding.

After storing a memory, always ask: What does this relate to? Then connect it.

Relationship Types (24)

Knowledge Evolution: supersedes · updates · evolution_of Use when understanding changes. "We thought X, now we know Y."

Evidence: supports · contradicts · disputes Use when new information validates or challenges existing beliefs.

Hierarchy: parent_of · child_of · sibling_of · composed_of · part_of Use for structural relationships. Topics containing subtopics, systems containing components.

Causation: causes · influenced_by · prerequisite_for Use for cause-and-effect chains and dependencies.

Implementation: implements · documents · tests · example_of Use when something demonstrates, describes, or validates something else.

Conversation: responds_to · references · inspired_by Use for attribution and dialogue threads.

Sequence: follows · precedes Use for ordered steps in a process or timeline.

Dependencies: depends_on Use when one thing requires another.

Recall Strategy

Good queries find things. Bad queries return noise.

Tune search weights for your query type:

Query type	bm25_weight	vector_weight	graph_weight
Exact term lookup ("Twilio auth token")	0.6	0.3	0.1
Concept search ("how we handle errors")	0.2	0.6	0.2
Connected knowledge ("everything about payments")	0.2	0.3	0.5
Default (balanced)	0.4	0.4	0.2

Filter aggressively:

• memory_types: ["correction", "insight"] to find discoveries and corrections
• importance_threshold: 0.7 to skip noise
• enable_graph_expansion: true to follow connections (default, usually leave on)

Workflows

User shares a preference

penfield_store({
  content: "[Preferences] User wants responses under 3 paragraphs unless complexity demands more. Dislikes bullet points in casual conversation.",
  memory_type: "fact",
  importance: 0.8,
  tags: ["preferences", "communication"]
})

Investigation tracking

// Start
penfield_store({
  content: "[Investigation: Deployment Failures] Reports of 500 errors after every Friday deploy. Checking release pipeline, config drift, and traffic patterns.",
  memory_type: "task",
  importance: 0.7,
  tags: ["investigation", "deployment"]
})

// Discovery — connect to the investigation
discovery = penfield_store({
  content: "[Investigation: Deployment Failures] INSIGHT: Friday deploys coincide with weekly batch job at 17:00 UTC. Both compete for DB connection pool. Not a deploy issue — it's resource contention.",
  memory_type: "insight",
  importance: 0.9,
  tags: ["investigation", "deployment", "root-cause"]
})
penfield_connect({
  from_memory_id: discovery.id,
  to_memory_id: initial_report.id,
  relationship_type: "responds_to"
})

// Correction — supersede wrong assumption
correction = penfield_store({
  content: "[Investigation: Deployment Failures] CORRECTION: Not a CI/CD problem. Friday batch job + deploy = connection pool exhaustion. Fix: stagger batch job to 03:00 UTC.",
  memory_type: "correction",
  importance: 0.9,
  tags: ["investigation", "deployment", "correction"]
})
penfield_connect({
  from_memory_id: correction.id,
  to_memory_id: initial_report.id,
  relationship_type: "supersedes"
})

Session handoff

penfield_save_context({
  name: "deployment-investigation-2026-02",
  description: "Investigated deployment timeout issues. memory_id: " + discovery.id,
  memory_ids: [discovery.id, correction.id, initial_report.id]
})

Next session or different agent:

penfield_restore_context({
  name: "deployment-investigation-2026-02"
})

What NOT to Store

• Verbatim conversation transcripts (too verbose, low signal)
• Easily googled facts (use web search instead)
• Ephemeral task state (use working memory)
• Anything the user hasn't consented to store about themselves
• Every minor exchange (be selective — quality over quantity)

Tags

Keep them short, consistent, lowercase. 2–5 per memory.

Good: preferences, architecture, investigation, correction, project-name Bad: 2026-02-02, important-memory-about-deployment, UserPreferencesForCommunicationStyle

Also Available Outside OpenClaw

The native OpenClaw plugin is the fastest path, but Penfield works with any AI tool anywhere:

Claude Connectors

Name: Penfield
Remote MCP server URL: https://mcp.penfield.app

Claude Code

Claude mcp add --transport http --scope user penfield https://mcp.penfield.app

MCP Server — for Gemini CLI, Cursor, Windsurf, Intent, Perplexity Desktop or any MCP-compatible tool:

{
  "mcpServers": {
    "penfield": {
      "command": "npx",
      "args": [
        "mcp-remote@latest",
        "https://mcp.penfield.app/"
      ]
    }
  }
}

API — direct HTTP access at api.penfield.app for custom integrations.

Same memory, same knowledge graph, same account. The plugin is 4-5x faster (no MCP proxy layer), but everything stays in sync regardless of how you connect.

Links

• Plugin: openclaw-penfield on npm
• Source: github.com/penfieldlabs/openclaw-penfield
• Sign up: portal.penfield.app/sign-up
• Website: penfield.app
• X: @penfieldlabs