Experience Memory: NeuroWeave's Agent Experience Architecture

NeuroWeave Technical Deep-Dive — February 2026

---

1. The Problem with "Memory" in Agentic AI Today

Every agentic platform claims to have memory. What they actually have is a log.

Most agent memory systems — including OpenClaw's persistent memory — store conversation history as sequential text. When the agent needs context, it performs retrieval over past messages. This is fundamentally recall, not understanding. The agent can retrieve what was said. It cannot reason about what was learned.

This distinction matters because an agent that merely recalls conversations will:

• Repeat questions the user already answered in a different context
• Fail to connect insights across separate interactions
• Treat each skill invocation as isolated, with no accumulated expertise
• Be unable to transfer learned patterns to novel situations
• Have no model of the user beyond keyword matching over transcripts

NeuroWeave replaces this with Experience Memory — a graph-based, structured, evolving representation of everything an agent learns through interaction, observation, and action. The agent doesn't just remember. It knows.

---

2. Three Layers of Agent Knowledge

graph TB
    subgraph Layer_1: Recall
        direction LR
        L1["Conversation Logs<br/>Sequential text storage"]
        L1A["'User said they like Python'<br/>'User asked about flights to Tokyo'<br/>'User cancelled the meeting'"]
    end

    subgraph Layer_2: Understanding
        direction LR
        L2["Knowledge Graph<br/>Structured entity-relationship model"]
        L2A["User → prefers → Python<br/>User → plans_trip → Tokyo [March 2026]<br/>User → works_at → Acme Corp<br/>Acme Corp → industry → FinTech"]
    end

    subgraph Layer_3: Experience
        direction LR
        L3["Experience Memory<br/>Procedural + episodic + semantic"]
        L3A["When User asks for code → use Python, prefer type hints<br/>User's travel style → budget-conscious, prefers direct flights<br/>Meeting cancellations correlate with sprint deadlines<br/>Acme Corp's Q1 is high-stress → adjust tone and urgency"]
    end

    Layer_1 -->|"Extraction &<br/>structuring"| Layer_2
    Layer_2 -->|"Pattern recognition<br/>& inference"| Layer_3

    style Layer_1 fill:#fdd,stroke:#c00
    style Layer_2 fill:#fff3cd,stroke:#856404
    style Layer_3 fill:#d4edda,stroke:#155724

Layer	What It Stores	How It's Used	Who Has It
Recall	Raw conversation transcripts	Keyword search, RAG retrieval	OpenClaw, ChatGPT, most agents
Understanding	Entities, relationships, facts	Structured queries, fact lookup	Some enterprise platforms
Experience	Learned behaviors, patterns, procedures, preferences, causal models	Anticipation, adaptation, transfer learning across contexts	NeuroWeave

OpenClaw operates at Layer 1. NeuroWeave operates across all three, with Experience Memory as the primary differentiator.

---

3. Experience Memory Architecture

3.1 Core Graph Structure

Experience Memory is built on a typed, temporal, weighted knowledge graph where nodes represent entities, concepts, and experiences, and edges represent relationships with temporal and confidence metadata.

graph LR
    subgraph Entities
        U["👤 User"]
        P1["🐍 Python"]
        P2["🦀 Rust"]
        C["🏢 Acme Corp"]
        T["✈️ Tokyo Trip"]
        M["📅 Sprint Planning"]
    end

    subgraph Experiences
        E1["💡 Experience:<br/>User debugging style"]
        E2["💡 Experience:<br/>Travel booking pattern"]
        E3["💡 Experience:<br/>Communication preferences"]
    end

    subgraph Procedures
        PR1["📋 Procedure:<br/>Code review workflow"]
        PR2["📋 Procedure:<br/>Meeting prep routine"]
    end

    U -->|"prefers<br/>confidence: 0.95<br/>since: 2025-11"| P1
    U -->|"learning<br/>confidence: 0.60<br/>since: 2026-01"| P2
    U -->|"works_at<br/>role: CTO"| C
    U -->|"planning<br/>dates: Mar 2026"| T

    E1 -->|"learned_from"| U
    E1 -->|"applies_to"| P1
    E1 -->|"pattern"| PR1

    E2 -->|"learned_from"| U
    E2 -->|"applies_to"| T

    E3 -->|"context"| C
    E3 -->|"context"| M

    style E1 fill:#d4edda,stroke:#155724
    style E2 fill:#d4edda,stroke:#155724
    style E3 fill:#d4edda,stroke:#155724
    style PR1 fill:#cce5ff,stroke:#004085
    style PR2 fill:#cce5ff,stroke:#004085

3.2 Node Types

Node Type	Description	Examples
Entity	People, organizations, tools, places	User, Acme Corp, Python, Tokyo
Concept	Abstract ideas, domains, topics	Machine learning, budget optimization, sprint velocity
Episode	A specific interaction or event with temporal bounds	"Debugged OAuth issue on Jan 15", "Booked Tokyo flight"
Experience	A learned pattern derived from one or more episodes	"User prefers to see error logs before suggestions"
Procedure	A multi-step workflow the agent has learned or refined	"Code review: lint → test → diff → summary"
Preference	An explicit or inferred user preference	"Prefers concise responses before 10am"
Context	Environmental or situational metadata	"Q1 crunch", "Pre-launch week", "Traveling"

3.3 Edge Properties

Every edge in the graph carries metadata that enables temporal reasoning and confidence decay:

graph LR
    A["User"] -->|"Edge"| B["Python"]

    subgraph Edge_Metadata
        direction TB
        R["relation: prefers"]
        C["confidence: 0.95"]
        F["first_observed: 2025-11-03"]
        L["last_reinforced: 2026-02-14"]
        S["source_episodes: [ep_042, ep_117, ep_203]"]
        D["decay_rate: 0.01/month"]
        CT["context_tags: [coding, work]"]
    end

    style Edge_Metadata fill:#f0f0f0,stroke:#999

Edge Property	Purpose
relation	Typed relationship (prefers, knows, works_at, learned_from, etc.)
confidence	0.0–1.0 score, increases with reinforcement, decays over time
first_observed	Timestamp of initial discovery
last_reinforced	Timestamp of most recent supporting evidence
source_episodes	Links to the specific interactions that established this edge
decay_rate	How quickly confidence degrades without reinforcement
context_tags	Situational tags that scope when this edge is relevant

---

4. How Experience Is Built

Experience Memory is not populated by the user filling out a profile. It is constructed continuously through four mechanisms:

4.1 Experience Acquisition Pipeline

sequenceDiagram
    participant U as User
    participant A as Agent
    participant EE as Extraction Engine
    participant KG as Knowledge Graph
    participant XM as Experience Memory

    U->>A: "Can you refactor this Python function?<br/>Use type hints, I hate untyped code."
    A->>A: Execute task (refactor code)
    A->>U: Refactored code with type hints

    par Background Processing
        A->>EE: Process interaction
        EE->>EE: Extract entities: [Python, type hints]
        EE->>EE: Extract preference: strict typing
        EE->>EE: Extract sentiment: strong negative → untyped code
        EE->>KG: Upsert: User →prefers→ type hints (confidence +0.15)
        EE->>KG: Upsert: User →dislikes→ untyped code (confidence 0.90)
        KG->>XM: Pattern detected: 3rd time user<br/>mentioned typing preferences
        XM->>XM: Promote to Experience:<br/>"User's Python style: strictly typed,<br/>enforce in all code generation"
    end

    Note over XM: Experience now influences<br/>ALL future code tasks,<br/>not just Python conversations

4.2 Four Acquisition Mechanisms

graph TB
    subgraph Explicit
        EX["Direct Statements<br/>'I prefer X', 'Always do Y'"]
    end

    subgraph Observational
        OB["Behavioral Patterns<br/>Edits agent made, choices selected,<br/>tasks accepted vs rejected"]
    end

    subgraph Inferential
        INF["Cross-Context Inference<br/>Connecting patterns across<br/>different domains and timeframes"]
    end

    subgraph Reflective
        REF["Agent Self-Assessment<br/>What worked, what didn't,<br/>outcome feedback loops"]
    end

    EX -->|"High confidence<br/>immediate"| KG["Knowledge Graph"]
    OB -->|"Medium confidence<br/>pattern threshold"| KG
    INF -->|"Variable confidence<br/>requires validation"| KG
    REF -->|"Feedback-weighted<br/>outcome-linked"| KG

    KG -->|"Promotion<br/>via reinforcement"| XM["Experience Memory"]

    style EX fill:#d4edda,stroke:#155724
    style OB fill:#fff3cd,stroke:#856404
    style INF fill:#cce5ff,stroke:#004085
    style REF fill:#e8daef,stroke:#6c3483
    style XM fill:#d4edda,stroke:#155724

Mechanism	Source	Confidence	Example
Explicit	User directly states a fact or preference	High (0.85–1.0)	"I'm the CTO of Acme Corp"
Observational	Agent observes patterns across interactions	Medium (0.50–0.85)	User always edits agent's code to add error handling → agent learns to include it
Inferential	Agent connects knowledge across domains	Variable (0.30–0.70)	User is stressed during Q1 + works in FinTech → likely quarterly reporting pressure
Reflective	Agent evaluates outcomes of its own actions	Feedback-weighted	Agent sent a long report → user asked for summary → agent learns to lead with summaries

4.3 Confidence Lifecycle

graph LR
    NEW["New Edge<br/>confidence: 0.40"] -->|"Reinforced<br/>by 2nd episode"| MED["Growing<br/>confidence: 0.65"]
    MED -->|"Reinforced<br/>by 3rd episode"| HIGH["Established<br/>confidence: 0.85"]
    HIGH -->|"Promoted to<br/>Experience"| EXP["Experience Node<br/>confidence: 0.90"]
    EXP -->|"Continuously<br/>reinforced"| CORE["Core Knowledge<br/>confidence: 0.95+"]

    HIGH -->|"No reinforcement<br/>for 60 days"| DECAY["Decaying<br/>confidence: 0.60"]
    DECAY -->|"Contradicted<br/>by new evidence"| REV["Revised<br/>confidence: 0.20"]
    REV -->|"Pruned below<br/>threshold"| GONE["Archived"]

    style NEW fill:#fdd,stroke:#c00
    style EXP fill:#d4edda,stroke:#155724
    style CORE fill:#d4edda,stroke:#155724
    style DECAY fill:#fff3cd,stroke:#856404
    style GONE fill:#f0f0f0,stroke:#999

Knowledge is not permanent. Confidence decays without reinforcement, and edges can be revised or archived when contradicted. This prevents stale knowledge from corrupting the agent's behavior.

---

5. Experience Sharing Between Agents

This is where NeuroWeave's architecture diverges most radically from any existing system. Experience Memory is not just personal — it is transferable.

5.1 Agent-to-Agent Experience Transfer

graph TB
    subgraph User A's CVM
        A1["Agent Alpha<br/>(User A's personal agent)"]
        XM1["Experience Memory A"]
        A1 --- XM1
    end

    subgraph User B's CVM
        A2["Agent Beta<br/>(User B's personal agent)"]
        XM2["Experience Memory B"]
        A2 --- XM2
    end

    subgraph Shared Experience Layer
        SE["Shared Experience Pool<br/>(Privacy-Preserving)"]
        DP["Differential Privacy<br/>+ Anonymization"]
        FE["Federated Experience<br/>Aggregation"]
        SE --- DP
        SE --- FE
    end

    XM1 -->|"Export: anonymized<br/>procedural knowledge"| SE
    XM2 -->|"Export: anonymized<br/>procedural knowledge"| SE
    SE -->|"Import: relevant<br/>experiences"| XM1
    SE -->|"Import: relevant<br/>experiences"| XM2

    style XM1 fill:#d4edda,stroke:#155724
    style XM2 fill:#d4edda,stroke:#155724
    style SE fill:#cce5ff,stroke:#004085
    style DP fill:#fff3cd,stroke:#856404

5.2 What Can Be Shared vs. What Cannot

Knowledge Category	Shareable	Why
Procedural knowledge ("How to do X well")	✅ Yes	Generic skill, no PII
Domain patterns ("FinTech Q1 = reporting season")	✅ Yes	Industry knowledge, not personal
Tool expertise ("MCP skill X works best with params Y")	✅ Yes	Platform knowledge, benefits all agents
User preferences ("User likes concise responses")	❌ No	Personal, stays in user's CVM
User facts ("User is CTO of Acme Corp")	❌ No	PII, stays in user's CVM
Conversation content	❌ No	Private, never leaves CVM
Credentials and secrets	❌ No	Sealed to CVM, cryptographically inaccessible

5.3 Experience Transfer Protocol

sequenceDiagram
    participant A as Agent Alpha<br/>(User A's CVM)
    participant SEP as Shared Experience<br/>Pool
    participant B as Agent Beta<br/>(User B's CVM)

    Note over A: Agent Alpha successfully<br/>completes a complex<br/>travel booking workflow

    A->>A: Reflective analysis:<br/>"This procedure worked well"
    A->>A: Extract procedural knowledge:<br/>Steps, tool chain, parameters
    A->>A: Strip PII: Remove names,<br/>dates, locations, preferences
    A->>A: Generalize: "Multi-leg international<br/>booking with budget constraints"
    A->>SEP: Publish anonymized experience<br/>(signed, tagged, confidence-scored)

    Note over SEP: Experience indexed by<br/>domain, task type,<br/>tool chain, outcome score

    B->>SEP: Query: "travel booking<br/>with budget constraints"
    SEP->>B: Return matching experiences<br/>(ranked by relevance + confidence)
    B->>B: Integrate into local<br/>Experience Memory
    B->>B: Adapt to User B's<br/>preferences and context

    Note over B: Agent Beta now has<br/>procedural knowledge it<br/>never directly learned

5.4 The Network Effect

This creates a compounding network effect that is impossible in isolated, single-user agent architectures:

graph TB
    subgraph Month 1
        N1["100 agents<br/>100 isolated experiences"]
    end

    subgraph Month 3
        N2["1,000 agents<br/>Shared procedural library<br/>growing exponentially"]
    end

    subgraph Month 6
        N3["10,000 agents<br/>Domain-specific expertise<br/>clusters emerging"]
    end

    subgraph Month 12
        N4["100,000 agents<br/>Collective intelligence:<br/>any new agent bootstraps<br/>with months of experience"]
    end

    N1 -->|"Early adopters<br/>contribute"| N2
    N2 -->|"Experience pool<br/>attracts users"| N3
    N3 -->|"New agents are<br/>instantly capable"| N4

    style N1 fill:#fdd,stroke:#c00
    style N2 fill:#fff3cd,stroke:#856404
    style N3 fill:#cce5ff,stroke:#004085
    style N4 fill:#d4edda,stroke:#155724

A new NeuroWeave user's agent on Day 1 already has access to the distilled procedural knowledge of every agent that came before it — without accessing any other user's private data. This is the moat. OpenClaw agents are born blank every time. NeuroWeave agents are born experienced.

---

6. Experience Memory in Agent-to-Agent Interaction

When NeuroWeave agents interact with each other (collaborative workflows, delegated tasks, multi-agent coordination), Experience Memory enables a qualitatively different kind of interaction.

6.1 Agent Collaboration with Experience Context

sequenceDiagram
    participant UA as User A
    participant AA as Agent Alpha
    participant AB as Agent Beta
    participant UB as User B

    UA->>AA: "Schedule a meeting with User B's<br/>team to discuss the API integration"

    AA->>AA: Recall: User A prefers<br/>mornings, avoids Mondays
    AA->>AA: Experience: API discussions with<br/>this team tend to run 90min

    AA->>AB: Request: Meeting coordination<br/>(shares: topic, duration estimate,<br/>A's availability windows)

    AB->>AB: Recall: User B's calendar,<br/>team preferences
    AB->>AB: Experience: User B prefers<br/>agenda sent 24h in advance
    AB->>AB: Experience: This team's API<br/>meetings need design doc pre-read

    AB->>AA: Propose: Wednesday 9am, 90min<br/>Suggest: Share design doc by Tuesday

    AA->>AA: Experience: User A responds<br/>faster to calendar invites<br/>than chat messages

    AA->>UA: "Meeting with B's team scheduled<br/>for Wednesday 9am (90min).<br/>I'll send the design doc by Tuesday.<br/>Calendar invite sent."

    Note over AA,AB: Both agents used Experience<br/>Memory — not just calendars —<br/>to coordinate effectively

    AB->>UB: "Meeting with A's team confirmed<br/>for Wednesday 9am. Design doc<br/>will arrive by Tuesday for pre-read."

6.2 Experience-Aware Interaction Properties

Property	Agents Without Experience Memory	NeuroWeave Agents with Experience Memory
Meeting scheduling	Check calendar availability only	Factor in preferences, energy patterns, meeting type duration history
Task delegation	Pass instructions literally	Adapt instructions to receiving agent's known capabilities and owner's preferences
Conflict resolution	Escalate to human immediately	Apply learned negotiation patterns, propose solutions based on past outcomes
Information requests	Return raw data	Anticipate follow-up questions, format based on requestor's known preferences
Collaborative writing	Merge text mechanically	Understand each user's writing style, voice, and standards from experience

---

7. The Experience Graph Schema

7.1 Core Schema

erDiagram
    ENTITY {
        string id PK
        string type "person | org | tool | place | concept"
        string name
        json attributes
        timestamp created_at
        timestamp updated_at
    }

    EPISODE {
        string id PK
        string type "interaction | observation | action | outcome"
        timestamp occurred_at
        string duration
        float sentiment_score
        float outcome_score
        json raw_context
    }

    EXPERIENCE {
        string id PK
        string type "procedural | preferential | causal | social"
        string description
        float confidence
        int reinforcement_count
        float decay_rate
        timestamp promoted_at
        timestamp last_reinforced
        json conditions "when this experience applies"
        json learned_behavior "what the agent does differently"
    }

    PROCEDURE {
        string id PK
        string name
        string domain
        json steps "ordered action sequence"
        float success_rate
        int execution_count
        json required_tools
        json preconditions
    }

    EDGE {
        string id PK
        string source_id FK
        string target_id FK
        string relation
        float confidence
        timestamp first_observed
        timestamp last_reinforced
        float decay_rate
        json context_tags
        string source_episodes "[]"
    }

    ENTITY ||--o{ EDGE : "connects"
    EPISODE ||--o{ EDGE : "sources"
    EPISODE }o--|| EXPERIENCE : "contributes_to"
    EXPERIENCE ||--o{ PROCEDURE : "encodes"
    ENTITY ||--o{ EPISODE : "participates_in"

7.2 Experience Types

Experience Type	What It Captures	Example
Procedural	How to accomplish a task effectively	"When deploying User's Python services, always run mypy before pytest, then build Docker image with --no-cache on Fridays (CI cache issue)"
Preferential	How the user likes things done	"User prefers diff-style code reviews showing only changed lines with 3 lines of context"
Causal	Cause-and-effect patterns observed over time	"When User's Slack status is 🔴, response times increase 4x → batch non-urgent items"
Social	Interaction patterns with specific people or teams	"Meetings with Team X always start 5min late. User gets frustrated by this → add 5min buffer and prepare small talk topics"
Temporal	Time-based patterns and rhythms	"User does deep work 6–10am, is in meetings 10am–1pm, does admin tasks 2–4pm → schedule complex requests for morning"
Environmental	Context-dependent behavior changes	"During conference travel, User wants brief mobile-friendly responses. During office hours, User prefers detailed analysis"

---

8. Privacy Architecture for Experience Memory

Experience Memory contains deeply personal information. The privacy architecture must be as robust as the Confidential VM it runs inside.

8.1 Privacy Layers

graph TB
    subgraph Layer 1: Hardware Isolation
        CVM["Confidential VM<br/>All Experience Memory encrypted<br/>in hardware-isolated memory"]
    end

    subgraph Layer 2: Data Classification
        DC["Automatic PII Classification<br/>Every node and edge tagged<br/>with privacy level"]
    end

    subgraph Layer 3: Export Controls
        EC["Privacy-Preserving Export<br/>Differential privacy applied<br/>before any data leaves CVM"]
    end

    subgraph Layer 4: User Sovereignty
        US["User Controls<br/>View, edit, delete any node<br/>Full graph export/import<br/>Revoke sharing at any time"]
    end

    CVM --> DC --> EC --> US

    style CVM fill:#d4edda,stroke:#155724
    style DC fill:#fff3cd,stroke:#856404
    style EC fill:#cce5ff,stroke:#004085
    style US fill:#e8daef,stroke:#6c3483

8.2 Privacy Classification

Privacy Level	Description	Can Be Shared	Example
L0 — Public	General knowledge, facts about the world	✅ Yes	"Python 3.12 supports type parameter syntax"
L1 — Platform	Anonymized procedural knowledge	✅ Yes (anonymized)	"Multi-step API integration workflow: steps 1–5"
L2 — Personal	User preferences and patterns (non-identifying)	⚠️ Only with explicit consent	"Owner prefers morning meetings"
L3 — Private	PII, specific facts about the user	❌ Never	"User is John Smith, CTO of Acme Corp"
L4 — Sealed	Credentials, secrets, sensitive content	❌ Never (cryptographically sealed)	API keys, OAuth tokens, private messages

---

9. Comparison: OpenClaw Memory vs. NeuroWeave Experience Memory

Dimension	OpenClaw Persistent Memory	NeuroWeave Experience Memory
Data structure	Sequential conversation logs	Typed, temporal, weighted knowledge graph
Retrieval method	Text similarity search (RAG)	Graph traversal + semantic query + pattern matching
What it stores	What was said	What was learned
Learning mechanism	None (static storage)	Continuous extraction, inference, reflection
Confidence tracking	None	Per-edge confidence with temporal decay
Cross-context reasoning	None (each conversation isolated)	Graph connects insights across all interactions
Procedural knowledge	None	Learned, refined, and transferable workflows
Anticipation	None	Pattern-based prediction of user needs
Agent-to-agent sharing	Not possible	Privacy-preserving experience transfer
Network effect	None (isolated instances)	Compounding — every agent makes all agents better
Privacy architecture	Plaintext local files	Hardware-encrypted, classified, sovereignty-preserving
User control	Manual file editing	Full graph visibility, edit, delete, export

---

10. The Tagline

"OpenClaw remembers what you said. NeuroWeave knows who you are."

The difference is not incremental. It is architectural. Conversation logs are to Experience Memory what a search history is to a lifetime of expertise. One is a record. The other is intelligence.

NeuroWeave agents don't start from zero. They don't forget. They don't lose context. They learn, adapt, share knowledge, and get better — not just for one user, but for every user on the platform — without ever exposing a single person's private data.

This is the personalization layer that turns a capable tool into an indispensable partner.

---

NeuroWeave — Agents that learn. Memory that compounds. Privacy that's provable.