Roadmap
This is a living document. Thought Fabric builds on top of flow-state-dev, so its timeline depends on the base framework's progress.
Shipped
Attention
filterRelevance— deterministic keyword-based relevance filtering (handler block)scoreSalience— LLM-based multi-dimension salience scoring (generator block)
Memory
- Working memory — bounded, salience-scored session resource with ACT-R power-law decay
workingMemoryCapture— bundled observe → remember → tick sequencer- Individual blocks: observe, remember, tick, snapshot, add
- Helpers: add, evict, pin, unpin, refresh, advance, items, formatForContext
workingMemoryContextFormatter— ready-made context slot for generators
Identity
perspective()— structured viewpoint model with config validation and frozen instancessystem()— factory that bundles blocks, capability, resources, and capture pipeline- Static blocks:
perspectiveApply,perspectiveAnalyze,perspectiveAuditor - Stateful blocks:
perspectiveObserve,perspectivePosition,perspectiveChallenge,perspectiveSnapshot,perspectiveAdvance - Resource-backed observations (session-scoped) and positions (configurable scope)
createPerspectiveCapability— capability withstaticandaccumulatedcontext presets- Helpers: add/remove/challenge/read/format for observations and positions
- Static formatters:
formatPerspective,summarizePerspective,perspectiveContextFormatter constitution()— ranked principle hierarchies with three conflict resolution modes (priority, weighted, contextual)constitutionAuditor— bundled sequencer: LLM review → deterministic enforce- Individual blocks:
constitutionReview(generator),constitutionEnforce(handler) - Helpers:
rankConstitutionPrinciples,computeConstitutionCompliance,formatConstitution,summarizeConstitutionReview
Metacognition (partial)
biasAnalyzer— bundled sequencer: detect → classify → score → counterpoint → format- Sycophancy detection with 4-dimension scoring (agreement, validating language, omitted counterpoints, framing adoption)
- 6 cognitive bias types: sycophancy, confirmation bias, anchoring, authority deference, recency bias, false consensus
- Counter-argument generation when score exceeds threshold
- Individual blocks exported for remixability:
biasDetectAgreement,biasClassify,biasScore,biasCounterpoint,biasFormat - Helper functions:
labelForSycophancyScore,computeCompositeSycophancyScore,shouldGenerateCounterpoints,summarizeBiasFindings - Conforms to
AnalyzerResultcontract for Response Auditor pattern integration
Coming Soon
Identity (remaining)
- Salience drift — perspective weights evolve with observation frequency
Perception
- Sensory processing and context framing
- Signal extraction from noisy input
- Multi-modal input normalization
Reasoning
- Structured deliberation patterns
- Chain-of-thought composition blocks
- Planning and strategy selection
Planned
Metacognition (remaining)
- Confidence calibration
- Strategy selection and self-correction
- Performance self-monitoring
Learning
- Pattern extraction from interaction history
- Skill acquisition and feedback integration
- Adaptive behavior modification
Design Principles
Each domain follows the same pattern:
- Blocks that compose with flow-state-dev sequencers and tools
- Helpers for direct resource manipulation
- Resources declared via
defineResource()with automatic flow integration - Context formatters for injecting domain state into LLM prompts
No separate runtime. No special execution model. Thought Fabric blocks are flow-state-dev blocks.