Memory Layers

Sci uses three memory layers modeled on cognitive science.

The three layers

Layer	Scope	Lifecycle	What lives here
Episodic	Profile-scoped	Accumulates; decay doesn't apply	Raw timestamped events, conversation fragments, decisions
Semantic	Profile-scoped	Promoted from episodic; Ebbinghaus decay	Durable facts, preferences, project context
Identity	Global (all profiles)	Stable; confidence-weighted	Values, skills, relationships — who you are

The profiles model

One person, multiple contexts. Think Safari browser profiles.

Identity (global)
  ├── preferences, values, cognitive style
  ├── relationships (exist across all profiles)
  └── Profiles
        ├── work    ← Claude Code, Cursor, Copilot
        ├── personal ← Claude.ai
        └── project:X ← (power users)

Default for new users: work + personal. Agents are assigned to a profile at connection time.

Episodic → Semantic promotion

Nightly consolidation lifts high-signal episodic memories into semantic nodes:

"I think we should use tsvector instead of dedicated search" (episodic)
                        ↓ nightly LLM pass
"Prefers tsvector for search — values operational simplicity" (semantic)

Promotion uses embedding similarity to deduplicate — if a similar semantic node already exists (cosine > 0.88), it gets reinforced instead of duplicated.

Ebbinghaus decay

Every semantic node has a decay_score recalculated nightly:

score = confidence × e^(−t / S)
  t = days since last access
  S = 1 + (access_count × 0.5)   (stability)

A note you've referenced 10 times has stability 6 — halved every ~6 days without access. A note you mentioned once has stability 1.5 — halved every ~1.5 days.

Nodes below 0.15 are flagged in metadata. They're not deleted — just deprioritized in retrieval.

The embeddings table

CREATE TABLE embeddings (
  memory_type  TEXT,     -- 'episodic', 'semantic', 'identity'
  memory_id    UUID,
  model_id     TEXT,     -- 'BAAI/bge-base-en-v1.5', 'voyage-3-lite', ...
  embedding    vector(768),
  UNIQUE (memory_type, memory_id, model_id)
);

Keyed by (type, id, model). This enables:

• Model migration without downtime — add rows for the new model alongside old ones
• A/B testing — compare retrieval quality between models on the same corpus
• Multiple dimensions — BGE (768-dim) and Voyage (512-dim) coexist

Retrieval

See Hybrid Retrieval for how recall searches across all three layers.