Namespace Modes

Namespaces isolate data and define default retrieval behaviour. Each namespace has a mode that tunes scoring weights, admission thresholds, traversal strategy, and compaction.

Choosing a mode

Mode comparison

	belief_system	agent_memory	general	procedural
Primary weight	Confidence (0.45)	Similarity (0.35)	Similarity (0.40)	Confidence + Similarity (0.40 each)
Decay	Slow (alpha=0.03)	Medium (0.05)	Medium (0.05)	Very slow (0.001)
Admission	Low bar (0.15)	Strict (0.35)	Balanced (0.25)	Strict (0.40)
Traversal	WaterCircle	Beam	WaterCircle	BFS
Max depth	4	2	3	3
Compaction	Off	RAPTOR	Off	Off
Best for	Chatbots, fact DBs	Agentic workflows	RAG, search	Skill storage

Usage

// Create or get a namespace with the desired mode
ns := db.Namespace("channel:general", namespace.ModeBeliefSystem)

// The mode sets defaults. You can override per-query
results, _ := ns.Retrieve(ctx, client.RetrieveRequest{
    Vector: queryVec,
    TopK:   10,
    // ScoreParams override is optional
})

belief_system

Designed for multi-source fact tracking where trust matters more than freshness.

• Source credibility is weighted highest (0.45)
• Troll and spam sources are rejected at the gate
• Wide graph traversal (depth 4) finds corroborating evidence
• Low admission threshold (0.15). Credibility gates retrieval, not ingestion

Use cases: Discord channel bots, knowledge bases with community input, fact-checking systems.

agent_memory

Designed for autonomous agents that learn from task outcomes.

• Utility feedback and recency are weighted heavily
• RAPTOR compaction automatically summarises similar memories
• Beam traversal for focused exploration
• Strict admission (0.35) avoids storing low-value episodes
• Memory types control decay: episodic fades fast, semantic persists longer

Use cases: LLM agents, task planners, conversational agents with long-term memory.

general

Balanced defaults suitable for most retrieval workloads.

• Similarity-first (0.40) for traditional vector search behaviour
• Moderate confidence and recency weighting
• Good starting point when you're not sure which mode to use

Use cases: RAG pipelines, document search, general-purpose retrieval.

procedural

Optimised for learned procedures and workflows that should persist.

• Very slow decay (alpha=0.001, half-life ~29 days)
• High confidence weight rewards validated procedures
• Strict admission (0.40) ensures only quality procedures are stored
• BFS traversal follows dependency chains

Use cases: Skill libraries, runbook storage, workflow automation.