Use-case profiles: specialized astrocyte subtypes

This document defines how the heterogeneity principle (Principle 4 from design-principles.md) is realized as configuration profiles - pre-built policy configurations tuned for specific agent use cases.

1. The heterogeneity principle

Biology: Astrocytes differ by region and lineage. Fibrous astrocyte in white matter behave differently from protoplasmic astrocyte in gray matter. Reactive phenotypes (A1/A2) are context-dependent, not one-size-fits-all.

Engineering: Different agent types need different memory governance. A customer support agent storing sensitive conversations has different needs than a coding assistant storing technical context. Rather than force every user to configure policies from scratch, Astrocyte ships profiles - named presets that configure the policy layer for common use cases.

2. Profile mechanism

A profile is a named configuration overlay that sets defaults for the policy layer. Users can extend or override any setting.

profile: support                   # Load the "support" profile as baseline
provider: mystique

# Override specific settings from the profile
homeostasis:
  recall_max_tokens: 8192          # Override the profile's default of 4096

Profiles are shipped as YAML files inside the astrocyte package. They are not code - they are configuration presets.

Resolution order: profile defaults → user config → per-bank overrides. Later values win.

3. Shipped profiles

3.1 `support` - Customer support agents

Optimized for: handling user conversations with PII sensitivity, empathetic recall, experience-heavy memory.

# profile: support
homeostasis:
  recall_max_tokens: 4096
  reflect_max_tokens: 4096
  retain_max_content_bytes: 51200
  rate_limits:
    retain_per_minute: 30
    recall_per_minute: 60
    reflect_per_minute: 10

barriers:
  pii:
    mode: regex
    action: redact
  validation:
    max_content_length: 10000
    reject_empty_content: true

signal_quality:
  dedup:
    enabled: true
    similarity_threshold: 0.92
  noisy_bank:
    enabled: true
    action: throttle

escalation:
  degraded_mode: empty_recall      # Agent can still function without memory

defaults:
  dispositions:
    skepticism: 2                  # Trust the customer's account
    literalism: 2                  # Flexible interpretation
    empathy: 5                     # High empathy
  preferred_fact_types:
    - experience                   # Prioritize what happened
    - world                        # Then general knowledge
  tags:
    - support
    - customer

Why these choices:

PII redaction is non-negotiable in support contexts.
High empathy disposition produces synthesis that acknowledges the user’s experience.
empty_recall degraded mode means the agent can still respond (just without personalization) when memory is down.
Lower rate limits because support conversations are human-paced.

3.2 `coding` - Coding assistants

Optimized for: storing technical context, code patterns, error resolutions, world-fact-heavy memory.

# profile: coding
homeostasis:
  recall_max_tokens: 8192          # Code context needs more tokens
  reflect_max_tokens: 8192
  retain_max_content_bytes: 204800 # 200KB - code files can be large
  rate_limits:
    retain_per_minute: 120         # High throughput for automated indexing
    recall_per_minute: 240
    reflect_per_minute: 30

barriers:
  pii:
    mode: regex
    action: warn                   # Warn but don't block - code has false positives
  validation:
    max_content_length: 100000
    reject_empty_content: true

signal_quality:
  dedup:
    enabled: true
    similarity_threshold: 0.98     # Higher threshold - similar code != duplicate
  noisy_bank:
    enabled: true
    action: warn

escalation:
  degraded_mode: error             # Code assistant should know memory is unavailable

defaults:
  dispositions:
    skepticism: 3                  # Neutral
    literalism: 5                  # High literalism - precision matters in code
    empathy: 1                     # Detached - technical context, not emotional
  preferred_fact_types:
    - world                        # Technical knowledge first
    - experience                   # Then project-specific learnings
  tags:
    - code
    - technical

Why these choices:

PII in warn mode because code contains email-like strings, regex patterns, and test data that trigger false positives.
Higher dedup threshold because similar code snippets are often meaningfully different.
High literalism disposition because code synthesis must be precise.
error degraded mode because a coding assistant giving wrong answers (without memory) is worse than saying “memory unavailable.”

3.3 `personal` - Personal assistants

Optimized for: broad context, temporal awareness, balanced dispositions, conversational memory.

# profile: personal
homeostasis:
  recall_max_tokens: 4096
  reflect_max_tokens: 8192
  retain_max_content_bytes: 51200
  rate_limits:
    retain_per_minute: 60
    recall_per_minute: 120
    reflect_per_minute: 30

barriers:
  pii:
    mode: regex
    action: redact
  validation:
    max_content_length: 20000
    reject_empty_content: true

signal_quality:
  dedup:
    enabled: true
    similarity_threshold: 0.93
  noisy_bank:
    enabled: true
    action: warn

escalation:
  degraded_mode: empty_recall

defaults:
  dispositions:
    skepticism: 3                  # Balanced
    literalism: 3                  # Balanced
    empathy: 4                     # Somewhat empathetic
  preferred_fact_types:
    - experience                   # User's personal experiences first
    - world
    - observation                  # Consolidated knowledge
  tags:
    - personal

3.4 `research` - Research and analysis agents

Optimized for: large document ingestion, high-volume recall, skeptical synthesis, graph-heavy retrieval.

# profile: research
homeostasis:
  recall_max_tokens: 16384         # Research needs deep context
  reflect_max_tokens: 16384
  retain_max_content_bytes: 524288 # 512KB - research documents are large
  rate_limits:
    retain_per_minute: 200         # Batch document ingestion
    recall_per_minute: 300
    reflect_per_minute: 50

barriers:
  pii:
    mode: disabled                 # Research data is typically non-personal
  validation:
    max_content_length: 500000
    reject_empty_content: true

signal_quality:
  dedup:
    enabled: true
    similarity_threshold: 0.97
  scoring:
    enabled: true                  # Quality-gate research inputs
    min_score: 0.4
    action: tag
  noisy_bank:
    enabled: true
    action: warn

escalation:
  degraded_mode: error             # Research needs accurate recall, not empty

defaults:
  dispositions:
    skepticism: 5                  # Highly skeptical - verify claims
    literalism: 4                  # Precise interpretation
    empathy: 1                     # Detached - analytical
  preferred_fact_types:
    - world                        # Factual knowledge
    - observation                  # Synthesized insights
  tags:
    - research
    - analysis

3.5 `minimal` - Bare minimum, no policies

For users who want the provider abstraction but no governance. Useful for testing, development, or when the provider handles everything internally.

# profile: minimal
homeostasis:
  recall_max_tokens: null          # No limit
  reflect_max_tokens: null
  retain_max_content_bytes: null
  rate_limits: null                # No rate limiting

barriers:
  pii:
    mode: disabled
  validation:
    reject_empty_content: true     # Only check: don't store nothing

signal_quality:
  dedup:
    enabled: false
  noisy_bank:
    enabled: false

escalation:
  degraded_mode: error

observability:
  otel_enabled: false
  prometheus_enabled: false

4. Custom profiles

Users can define their own profiles as YAML files and reference them by path:

profile: ./profiles/healthcare-agent.yaml
provider: mystique

Or inline, by omitting the profile key and configuring everything directly.

5. Per-bank profile overrides

A single Astrocyte instance can serve multiple banks with different profiles. This supports multi-tenant applications where each tenant (or agent) has different memory governance needs:

profile: support                   # Default for all banks
provider: mystique

banks:
  vip-customer:
    profile: support               # Explicit (same as default here)
    barriers:
      pii:
        mode: llm                  # Upgrade to LLM-based PII for VIP
        action: reject             # Stricter: reject, don't redact

  internal-testing:
    profile: minimal               # Different profile entirely

  research-team:
    profile: research
    homeostasis:
      recall_max_tokens: 32768     # Research team gets bigger budget

6. Profile selection guidance

Agent type	Profile	Key tradeoff
Customer support, helpdesk	`support`	PII protection vs. recall fidelity
IDE assistant, code review	`coding`	Throughput vs. false-positive PII
Personal assistant, chat	`personal`	Breadth vs. privacy
Document analysis, RAG	`research`	Depth vs. cost (large token budgets)
Development, testing	`minimal`	Speed vs. safety

Profiles are starting points. Production deployments should audit and tune settings based on observed metrics (see policy-layer.md, section 5: Observability).