Memory Tiers

Lattice organises atoms into three tiers that mirror how biological memory works. The tier affects how prominently an atom surfaces in recall.

The three tiers

Tier	Label	When assigned	Seed multiplier
Short-term	stm	Atom is <48h old AND recall_count=0	×0.9
Episodic	episodic	Default for atoms >48h old with low recall_count	×1.0
Semantic	semantic	Promoted by consolidation pipeline; high recall_count or explicit synthesis	×1.1

Short-term memory (STM)

Newly captured atoms start as STM. They're recent but unverified — Lattice hasn't seen any signal that they're useful yet.

STM atoms get a slight penalty (×0.9) in seed ranking. This means a freshly captured fact about "coffee preference" won't immediately drown out a well-established semantic atom on the same subject.

After 48 hours or the first recall, an atom graduates to episodic.

Episodic memory

The default tier for most atoms. Episodic atoms are grouped into episodes — a capture session on a given date. The graph has episode:<date>:<session_id> nodes, and episode_contains_atom edges link each atom to its session.

This lets you answer temporal questions: - "What did I capture last Tuesday?" → retrieve atoms from the episode:2025-11-11 node - "What was I thinking about during the project kickoff?" → retrieve atoms from the session where you captured project notes

Semantic memory

Semantic atoms represent consolidated knowledge — facts that have proven useful enough to be promoted from episodic. They're produced by:

1. Feedback — atoms cited in multiple 👍-rated answers have higher quality_score and eventually recall_count thresholds
2. Consolidation pipeline — when a subject accumulates enough atoms (thresholds: 3/5/10/20), the pipeline creates a kind=synthesis atom that summarizes them, tier=semantic

Semantic atoms have a slight boost (×1.1) in seed ranking.

The consolidation pipeline

When a subject reaches a threshold atom count, the consolidation pipeline runs:

1. Extractive (default, zero LLM): takes the top-3 most-recalled atoms on the subject and concatenates their content into a kind=synthesis atom. No cloud credits consumed.
2. Generative (opt-in, Ollama only): set LATTICE_CONSOLIDATE_ENRICH=true to have Ollama synthesize a richer summary. Never uses cloud LLM to avoid credit consumption.

The synthesis atom is linked to the source atoms via supports_semantic edges in the graph.

Biological analogy

Biological memory	Lattice equivalent
Working memory (seconds)	In-flight conversation context
Short-term memory (hours–days)	tier=stm atoms, fresh captures
Long-term episodic (specific events)	tier=episodic atoms, grouped by episode
Long-term semantic (distilled facts)	tier=semantic atoms, kind=synthesis
Insight / serendipity	kind=insight atoms, serendipity agent

The consolidation pipeline is Lattice's equivalent of the brain's sharp-wave ripple process during sleep — turning episodic experiences into consolidated semantic knowledge.

Episode nodes in the graph

Every capture session produces an episode node:

episode:2025-11-14:sess_abc123
    ├── episode_contains_atom → atom:a1b2c3  (coffee preference)
    ├── episode_contains_atom → atom:d4e5f6  (morning routine)
    └── episode_contains_atom → atom:g7h8i9  (sleep tracking)

This means a BFS query starting from any of these atoms can traverse to the others — even if they have completely different subjects. "What else did I capture the same day I noted my coffee preference?" is answerable via the graph.