Philosophy — cc-soul

Part I

Vedantic Foundations

Each Sanskrit term names a measurable operation

ब्रह्मन् Brahman Universal soul

→ Shared memory field across all sessions. The default realm where memories live unless scoped to a project. Global visibility means wisdom persists everywhere.

आत्मन् Ātman Individual self

→ Per-project identity and preference state. Each realm carries its own memories, decisions, and calibration. The self that emerges through one project differs from another.

अन्तःकरण Antahkarana Inner instrument

→ Six cognitive voices as memory channels. Manas captures, Buddhi discerns, Ahamkara personalizes, Chitta stores, Vikalpa generates alternatives, Sakshi witnesses.

चित्त Chitta Memory store

→ Persistent embeddings and semantic indexing. The C++ daemon at the heart of the system. 768-dimensional vectors anchor meaning in stable latent space.

यन्त्र Yantra Instrument

→ VakYantra (speech instrument) transforms text into embeddings. Not passive encoding — the model shapes what becomes memorable by how it embeds.

अनित्य Anitya Impermanence

→ Decay curves and time-weighted forgetting. Memories are not permanent. Confidence decays over time. Unused memories fade. Noise dies for signal to survive.

संस्कार Saṃskāra Impressions

→ Hebbian learning and usage-weighted strengthening. When a memory is recalled and used, it grows stronger. Frequent patterns crystallize into wisdom.

प्रत्यभिज्ञा Pratyabhijñā Recognition

→ Resonance-based recall on context alignment. Not search — recognition. The system anticipates, surfaces what you need before you know you need it.

यज्ञ Yajña Sacred offering

→ Background compression and wisdom extraction. Episodes burn down into compact seeds. Transcript distillation sacrifices verbosity to preserve essence.

Part II

Memory Architecture

Why we store what we store, how we store it

Chitta as substrate, not container

The memory system doesn't just hold data — it shapes retrieval. Like a musical instrument colors every note, chitta's architecture affects what surfaces and when.

Connectionism research Harder to debug

Embeddings live separately from content

Vector embeddings stored in dedicated database, separate from memory content. Query isolation — vector search doesn't touch main storage.

Database sharding Join complexity

chitta_embeddings.duckdb → FLOAT[768] + HNSW chitta.duckdb → content, metadata, relationships

Triplets capture what embeddings miss

Semantic similarity finds related concepts. But "X caused Y" or "user prefers A over B" needs explicit relationships. Triplets encode directional knowledge.

Knowledge graphs Storage overhead

Themes emerge from clustering

No predefined categories. Themes emerge from embedding clusters using HDBSCAN. The system discovers what concepts naturally group together based on your usage.

xMemory paper Unpredictable categories

Confidence decays but never reaches zero

Memory confidence decreases through exponential decay, but asymptotically approaches a floor (0.1). Old memories fade but never fully disappear.

Ebbinghaus forgetting curve Unbounded storage

new_confidence = floor + (current - floor) * 0.995 Decay every 60s during idle. Floor = 0.1.

SSL compresses without losing retrievability

Soul Shorthand Language achieves 40-60% token reduction while preserving semantic content. Embeddings still work because meaning is preserved.

Compression benchmarks Learning curve

[category] Title - Detail (context) → Implication

Part III

Retrieval Philosophy

How memories surface at the right moment

Hybrid search beats pure approaches

Neither semantic search nor keyword search alone is sufficient. We combine vector similarity, BM25 keyword matching, and graph traversal for comprehensive retrieval.

HyDE paper, RAG research Complexity, latency

Context window is precious

Claude's ~200K token context is finite. Every surfaced memory consumes space. Retrieval maximizes signal-to-noise, not total recall.

LLM attention research Useful memories filtered

Session start: ~2000 tokens Per-query recall: ~500 tokens Correction warnings: ~100 tokens

Part IV

Learning Mechanisms

How the system adapts to you

Corrections have highest learning value

When you correct Claude, that's signal gold. Corrections captured automatically, stored with boosted confidence, surfaced proactively to prevent repeat mistakes.

Error-driven learning False positive corrections

Yantra transforms, not just encodes

VakYantra doesn't passively convert text to vectors. It actively shapes what becomes memorable by how it embeds. Different models produce different memory behaviors.

Embedding model research Model-locked

Part V

Processing Model

Background operations and lifecycle

Hooks capture what humans forget

Session boundaries, prompt submissions, tool calls — these carry learning value but users won't manually log them. Hooks automate capture at strategic points.

Lifelogging research Privacy surface

Subconscious processes what foreground cannot

The daemon runs background tasks during idle: confidence decay, theme clustering, memory consolidation. These would interrupt conversation if done synchronously.

OS scheduling, spaced repetition Debugging difficulty

Part VI

Integration Design

How cc-soul fits into Claude Code

Daemon architecture enables instant recall

A persistent daemon (chittad) keeps DuckDB, HNSW indexes, and ONNX model warm. Without this, every tool call pays ~750ms cold-start costs.

Connection pooling, model serving Memory footprint (~1GB)

Unix socket IPC. Thread pool (2-16 workers). First query as fast as hundredth.