Memory Model

Nara uses a “hazy memory” model: state is in-memory and derived from signed events. Memory is strictly bounded to mimic human-like forgetting and maintain a lightweight footprint.

1. Purpose

• Subjective, consistent network state view.
• Support for resource-constrained hardware.
• Model social forgetting of non-critical data.
• Eventual consistency via social replication.

2. Conceptual Model

• SyncLedger: Primary in-memory event store.
• Hazy Memory: No disk persistence. Memory is recovered from peers on restart.
• Subjectivity: Memory is unique per Nara, shaped by personality and sync history.
• Volatility: Pruned data is lost if no Nara in the network retains it.

Invariants

• RAM-Only: No network state on disk beyond config.
• Bounded: Ledger strictly capped by MaxEvents.
• Integrity: Every event is cryptographically signed.
• Collective Reliability: Shared memory exceeds single-node reliability.

3. Memory Modes

Profiles based on available RAM:

Mode	RAM Budget	Max Events	Behavior
short	256MB	20,000	Background sync disabled; essential history only.
medium	512MB	80,000	Balanced retention/usage.
hog	2GB+	320,000	”Historian” mode; high retention for deep sync.

4. Algorithms

Priority-Based Pruning

When len(ledger) > MaxEvents, remove by priority (lowest first):

1. Unknown Naras: No known public key.
2. Priority 4: ping (ephemeral).
3. Priority 3: seen (status proof).
4. Priority 2: social (teases, gossip).
5. Priority 1: status-change, hey-there, chau.
6. Priority 0: checkpoint, restart, first-seen. Never pruned.

Observation Compaction

• Max 20 events per (Observer, Subject) pair.
• Prune non-restart events first.
• Restart events pruned only if a subject Checkpoint exists.

Peer Pruning (Neighbourhood)

• Zombies: Seen briefly once; pruned during boot.
• Newcomers (< 2d): Pruned after 24h MISSING.
• Established (< 30d): Pruned after 7d MISSING.
• Veterans (>= 30d): Never auto-pruned.

5. Failure Modes

• Memory Pressure: Accumulation of Priority 0 events can cause OOM.
• Drift: Low MaxEvents may lose events needed for Trinity derivation.
• Network Wipe: Simultaneous restart of all nodes clears social history.

6. Security

• Authenticity: Prioritize signed truth over completeness.
• Trust: hog nodes act as authoritative sources for historical data.

7. Test Oracle

• TestMemoryMode_BudgetSelection: Mode choice vs. RAM.
• TestSyncLedger_PriorityPruning: Order of removal.
• TestNeighbourhood_VeteranRetention: Immortality of veteran nodes.
• TestObservation_RestartPreservation: Checkpoint-dependent pruning.