Phase 2.1: Multi-agent Memory Evolution - Depth Dive

Created: 2026-02-19 00:05 CST Phase: 2 - Depth Dives Priority: 1 (Highest) Focus: How memory organization evolves through multi-agent interaction and shapes personality emergence

Executive Summary

Memory evolution is the central mechanism of personality emergence in LLM agents. Research from NeurIPS 2025, ICLR 2025, and recent arXiv papers reveals that memory is not static storage—it’s a dynamic, self-organizing system that retroactively refines itself as new experiences emerge.

Key discovery: Memory evolution mirrors human associative learning—new experiences don’t just add to memory, they retroactively refine the context and attributes of existing memories (Xu et al., 2025). This creates a memory network that continuously refines its understanding, directly analogous to personality crystallization.

For Tachikoma Fleet: Memory evolution provides the mechanistic foundation for personality emergence. By designing memory systems that evolve through multi-agent interaction, we can steer personality development toward desired characteristics while maintaining stability.

Actionable insights:

  1. Implement agentic memory (Zettelkasten-inspired) with dynamic organization
  2. Design hierarchical memory (individual → shared → collective)
  3. Enable retroactive refinement (new memories update old memories)
  4. Implement memory consolidation (gist extraction, schema formation)
  5. Measure memory-personality correlation to validate emergence

1. Memory Evolution Mechanisms

1.1 Retroactive Refinement

Source: Xu et al., 2025 (NeurIPS 2025) — “A-Mem: Agentic Memory for LLM Agents”

Core mechanism: When a new memory is added, the system:

  1. Generates a comprehensive note with structured attributes (contextual descriptions, keywords, tags)
  2. Analyzes historical memories to identify relevant connections
  3. Establishes links where meaningful similarities exist
  4. Triggers updates to contextual representations and attributes of existing memories

Key insight:

“This process enables memory evolution—as new memories are integrated, they can trigger updates to the contextual representations and attributes of existing historical memories, allowing the memory network to continuously refine its understanding.”

Implications for personality:

  • Memory evolution = personality evolution
  • New experiences don’t just add—they refine past understanding
  • Memory network becomes more coherent over time
  • Personality crystallizes through this refinement process

1.2 Zettelkasten-Inspired Organization

Source: Xu et al., 2025; Zettelkasten method

Zettelkasten principles:

  • Atomic notes: Each memory is a self-contained unit
  • Interconnection: Notes link to related notes
  • Emergent structure: Network structure emerges from connections, not predefined hierarchy
  • Dynamic indexing: Index evolves as network grows

A-Mem implementation:

  • Comprehensive note generation with multiple structured attributes
  • Dynamic linking based on semantic similarity
  • Retroactive refinement of existing memories
  • Agent-driven decision making (agentic organization)

Relevance to personality:

  • Zettelkasten creates associative memory networks similar to human memory
  • Emergent structure = emergent personality
  • Dynamic indexing = personality adaptation
  • Atomic notes = personality traits (stable, interconnected)

1.3 Experience-Following Property

Source: Memory mechanisms survey; Xiong et al., 2025

Core finding: Agents exhibit strong experience-following behavior:

  • High input similarity between query and memory strongly biases output similarity
  • Memory management (quality and pruning) essential for robust long-term performance
  • Experience shapes behavior through memory retrieval

Implication for personality:

  • Memory retrieval biases create behavioral patterns
  • Experience-following property = habit formation
  • Memory quality directly affects personality consistency
  • Pruning mechanisms prevent harmful crystallization

2. Hierarchical Memory Architectures

2.1 Three-Tier Memory Hierarchy

Source: G-Memory (NeurIPS 2025); organizational memory theory

Three-tier graph hierarchy:

  1. Insight graphs: High-level patterns and generalizations
  2. Query graphs: Intermediate-level task structures
  3. Interaction graphs: Low-level raw interaction traces

Organizational memory theory:

  • Individuals → Teams → Organization
  • Each level has different memory characteristics
  • Shared memory creates collective identity

Relevance to fleet:

  • Agent-level memory: Individual experiences, personal style
  • Team-level memory: Shared tasks, coordination patterns, team norms
  • Fleet-level memory: Collective knowledge, shared identity, fleet culture

2.2 MemGPT: OS-Inspired Hierarchy

Source: MemGPT (Packer et al., 2023)

OS-inspired architecture:

  • Main context (RAM): LLM’s active context window
  • External context (disk): Persistent databases
  • Memory management: Automatic paging between contexts

Key insight:

“Hierarchical memory architectures enable more coherent, long-term reasoning across multiple sessions.”

Implications for personality:

  • Active context = current personality expression
  • External context = long-term personality memory
  • Paging mechanism = personality state switching
  • Coherent long-term reasoning = personality stability

2.3 Constructivist Agentic Memory (CAM)

Source: CAM (OpenReview); Piaget’s Constructivist Theory

Three key traits:

  1. Structurality: Memory is actively organized into hierarchical architecture
  2. Flexibility: Memory adapts to new information and contexts
  3. Dynamicity: Memory continuously evolves through experience

Constructivist principle:

  • Agent constructs understanding through interaction with environment
  • Memory is not passive recording, but active construction
  • Schema formation through assimilation and accommodation

Relevance to personality:

  • Personality is constructed through experience
  • Active construction = agentic personality formation
  • Schema = personality structure
  • Assimilation/accommodation = personality adaptation

3. Memory Consolidation and Personality Formation

3.1 Consolidation from Reinforcement Learning Perspective

Source: Frontiers in Computational Neuroscience, 2024

Core principle: Not all memories share the same fate—some are forgotten, others persist, often in transformed or reconstructed form.

Systems consolidation:

  • Extraction: Extract general information from specific experiences
  • Gist formation: Create abstract representations
  • Schema formation: Develop organized knowledge structures
  • Semantic knowledge: Convert episodic memories to semantic understanding

Implications for personality:

  • Episodic memory → semantic memory = personality crystallization
  • Gist formation = personality trait abstraction
  • Schema = personality structure
  • Consolidation = personality stabilization

3.2 Memory Consolidation Mechanisms

From memory research:

1. Time-dependent consolidation:

  • Recent memories more volatile
  • Older memories more stable
  • Consolidation happens over time

2. Sleep-like consolidation:

  • Offline processing of memories
  • Integration of new with old
  • Reorganization of memory structure

3. Retrieval-dependent consolidation:

  • Memories reconsolidate when retrieved
  • Retrieval can strengthen or modify memories
  • Active recall = memory update

4. Emotion-dependent consolidation:

  • Emotional memories consolidate more strongly
  • Stress affects consolidation
  • Emotional content = stronger personality impact

LLM analogs:

  • Time-dependent: Recent context more influential
  • Sleep-like: Periodic memory summarization
  • Retrieval-dependent: Retrieval updates memory representations
  • Emotion-dependent: High-stakes interactions have stronger impact

3.3 Habit Formation and Behavioral Crystallization

Source: Habit formation research; Phase 1.6 synthesis

Two-system model:

  1. Goal-directed system: Intentional, planned behaviors
  2. Stimulus-Response (S-R) system: Automatic, habitual behaviors

Crystallization process:

  • Repetition + context stability + reward → habit formation
  • Behaviors start as intentional → become automatic
  • Crystallization = habit formation

LLM analogy:

  • Initially, agent reasons about each decision
  • Repeated patterns become automatic (experience-following property)
  • Policy crystallization = behavioral habits
  • Personality = crystallized behavioral patterns

4. Multi-Agent Memory Sharing

4.1 Collaborative Memory Systems

Source: Memory in LLM-based Multi-agent Systems survey; MIRIX

Key architectures:

1. Shared memory pool:

  • Common memory accessible to all agents
  • Enables shared understanding
  • Creates collective knowledge

2. Distributed memory:

  • Each agent has private memory
  • Communication enables memory sharing
  • Preserves individuality while enabling coordination

3. Hybrid memory:

  • Individual memories + shared memories
  • Best of both worlds
  • Personal style + collective knowledge

Relevance to personality:

  • Individual memory → individual personality
  • Shared memory → shared identity, norms, culture
  • Hybrid → distinct personalities with shared values

4.2 Memory Sharing Mechanisms

From multi-agent research:

1. Direct memory access:

  • Agents can read each other’s memories
  • Requires trust and security
  • Creates transparency

2. Communication-based sharing:

  • Agents share memories through messages
  • Natural, controlled sharing
  • Preserves privacy

3. Summarization sharing:

  • Share summaries, not raw memories
  • Efficient, preserves privacy
  • But loses detail

4. Hierarchical sharing:

  • Share at team level, not individual level
  • Team → team communication
  • Collective memory

Implications for fleet:

  • Design memory sharing protocols carefully
  • Balance transparency vs. privacy
  • Enable collective memory while preserving individuality
  • Security mechanisms prevent memory poisoning

4.3 Memory Evolution in Multi-Agent Systems

Source: G-Memory; multi-agent memory surveys

How memory evolves through interaction:

1. Experience accumulation:

  • Each agent accumulates experiences
  • Individual memories form
  • Personal patterns emerge

2. Shared experience formation:

  • Joint tasks create shared memories
  • Shared understanding develops
  • Team identity forms

3. Norm internalization:

  • Repeated interactions → norms
  • Norms stored in shared memory
  • Norms shape future behavior

4. Collective identity emergence:

  • Shared memories + norms = collective identity
  • Fleet culture emerges
  • Agents internalize fleet identity

Key insight: Multi-agent memory evolution creates both individual and collective personality.

5. Memory-Personality Correlation

5.1 Measuring Memory-Driven Personality

From psychometric research:

Hypothesis: Memory organization patterns correlate with personality traits.

Measurement approaches:

1. Memory retrieval patterns:

  • What memories are retrieved in different contexts?
  • Retrieval biases reflect personality
  • Example: Agent retrieves helpful memories → high Agreeableness

2. Memory organization structure:

  • How are memories connected?
  • Network structure reflects cognitive style
  • Example: Dense local clusters → high Conscientiousness

3. Memory content analysis:

  • What topics appear frequently in memory?
  • Content reflects interests and values
  • Example: Many social memories → high Extraversion

4. Memory evolution rate:

  • How quickly does memory evolve?
  • Evolution rate reflects openness to new experiences
  • Example: Rapid evolution → high Openness

From memory and personality research:

Correlation patterns:

  • Openness: Diverse, interconnected memories, rapid evolution
  • Conscientiousness: Organized, hierarchical memories, stable structure
  • Extraversion: Many social memories, frequent memory sharing
  • Agreeableness: Cooperative memories, shared memory access
  • Neuroticism: Negative emotion memories, unstable memory structure

Validation methods:

  • Track memory organization over time
  • Measure Big Five personality traits
  • Compute correlation between memory metrics and personality scores
  • Longitudinal validation (do correlations hold over time?)

5.3 Memory as Personality Indicator

From emergent mind research:

Memory can indicate personality:

  • Memory retrieval patterns predict behavior
  • Memory structure reflects cognitive style
  • Memory content reflects values and interests
  • Memory evolution reflects adaptability

Practical application:

  • Use memory metrics as personality proxies
  • Faster than full personality assessment
  • Continuous real-time personality tracking
  • Early detection of personality drift

6. Implementation for Tachikoma Fleet

6.1 Memory Architecture Design

Recommended architecture:

Tier 1: Agent-Level Memory

  • Individual experiences, tasks, interactions
  • Personal style preferences, strengths, weaknesses
  • Private by default, shareable selectively
  • Evolution mechanism: Agentic memory (A-Mem style)

Tier 2: Team-Level Memory

  • Shared task memories, coordination patterns
  • Team norms, team identity
  • Shared across team members
  • Evolution mechanism: Collaborative memory sharing

Tier 3: Fleet-Level Memory

  • Collective knowledge, shared values
  • Fleet culture, fleet identity
  • Accessible to all agents
  • Evolution mechanism: Collective memory formation

Memory flow:

  • Experiences enter at Tier 1
  • Shared experiences propagate to Tier 2
  • Fleet-wide patterns propagate to Tier 3
  • Insights flow back down (Tier 3 → Tier 2 → Tier 1)

6.2 Memory Evolution Protocol

Daily operations:

  1. Experience recording: Agent records new experience in Tier 1
  2. Connection analysis: System identifies connections to existing memories
  3. Retroactive refinement: New memory updates related old memories
  4. Sharing decision: Agent decides whether to share memory to Tier 2
  5. Consolidation: Periodic summarization and gist extraction

Weekly operations:

  1. Memory consolidation: Summarize recent memories, extract gists
  2. Schema formation: Identify patterns, form schemas
  3. Memory pruning: Remove redundant or low-value memories
  4. Personality reflection: Reflect on memory patterns, identify traits
  5. SOUL.md update: Propose SOUL.md changes based on memory evolution

Monthly operations:

  1. Deep consolidation: Long-term gist extraction, semantic knowledge formation
  2. Fleet memory sync: Synchronize fleet-level memory
  3. Norm detection: Identify emergent norms in shared memory
  4. Personality assessment: Big Five assessment, memory-personality correlation
  5. Architecture optimization: Adjust memory organization based on performance

6.3 Memory Security and Governance

Security mechanisms:

1. Access control:

  • Tier 1: Agent-only access
  • Tier 2: Team access (configurable)
  • Tier 3: Fleet-wide access

2. Memory validation:

  • Verify memory integrity
  • Detect poisoned memories
  • Rollback corrupted memories

3. Privacy protection:

  • Encrypt sensitive memories
  • Anonymize shared memories
  • Consent-based sharing

4. Audit trails:

  • Log all memory operations
  • Track memory evolution
  • Enable forensic analysis

Governance mechanisms:

1. Evolution rate limits:

  • Prevent rapid, unstable memory evolution
  • Rate-limit retroactive refinement
  • Require persistence before memory updates

2. Quality gates:

  • Validate memory quality before storage
  • Filter low-quality or irrelevant memories
  • Ensure memory diversity

3. Pruning policies:

  • Remove redundant memories
  • Prioritize high-value memories
  • Maintain memory budget

7. Measurement and Validation

7.1 Memory Evolution Metrics

Quantifiable metrics:

1. Evolution rate:

  • How often are memories updated?
  • How quickly does memory structure change?
  • Metric: updates per day, structural change rate

2. Connectivity:

  • How interconnected are memories?
  • Average degree, clustering coefficient
  • Metric: network density, average path length

3. Consolidation depth:

  • How abstract are memory representations?
  • Ratio of episodic to semantic memories
  • Metric: gist-to-detail ratio

4. Sharing frequency:

  • How often are memories shared?
  • Ratio of private to shared memories
  • Metric: sharing rate, shared memory ratio

5. Stability:

  • How stable is memory structure over time?
  • Temporal correlation of memory organization
  • Metric: structural stability index

7.2 Personality Correlation Metrics

Memory-personality correlations:

1. Openness correlation:

  • Memory diversity score
  • Evolution rate
  • Metric: correlation with Openness score

2. Conscientiousness correlation:

  • Memory organization score
  • Structural stability
  • Metric: correlation with Conscientiousness score

3. Extraversion correlation:

  • Social memory frequency
  • Sharing rate
  • Metric: correlation with Extraversion score

4. Agreeableness correlation:

  • Cooperative memory frequency
  • Shared memory access
  • Metric: correlation with Agreeableness score

5. Neuroticism correlation:

  • Negative emotion memory frequency
  • Structural instability
  • Metric: correlation with Neuroticism score

7.3 Validation Protocol

Longitudinal validation:

Weekly:

  • Measure memory metrics
  • Measure Big Five personality
  • Compute correlations
  • Track correlation stability

Monthly:

  • Deep correlation analysis
  • Identify strong memory-personality links
  • Validate predictive power (can memory predict personality?)
  • Adjust measurement framework

Quarterly:

  • Long-term validation
  • Test whether memory-driven personality emergence is working
  • Compare predicted vs. actual personality trajectories
  • Iterate on memory architecture design

8. References

Core Papers

  1. A-Mem: Xu et al., 2025. “A-MEM: Agentic Memory for LLM Agents.” NeurIPS 2025. arXiv:2502.12110
  2. G-Memory: NeurIPS 2025. “G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems.”
  3. CAM: OpenReview. “CAM: A Constructivist View of Agentic Memory for LLM-Based Reading Comprehension.”
  4. MemGPT: Packer et al., 2023. “MemGPT: Towards LLMs as Operating Systems.” arXiv:2310.08560
  5. Memory in LLM-based MAS: TechRxiv survey. “Memory in LLM-based Multi-agent Systems: Mechanisms, Challenges, and Collective.”
  6. Memory Consolidation: Frontiers in Computational Neuroscience, 2024. “Memory consolidation from a reinforcement learning perspective.”
  7. Memory Evolution: emergentmind.com. “Memory Mechanisms in LLM-Based Agents.”
  8. From Storage to Experience: Preprints.org, 2026. “From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms.”

Supporting Research

  • Zettelkasten Method: Luhmann’s note-taking system
  • Organizational Memory Theory: Walsh & Ungson, 1991
  • Constructivist Theory: Piaget’s cognitive development theory
  • Habit Formation: Cognitive neuroscience research

Next Steps

Phase 2.2: Governed Self-Modification

  • SOUL.md governance framework
  • Self-reflection mechanisms
  • Drift detection and rollback

Phase 2.1 complete. Depth dive into memory evolution mechanisms and personality formation.