Profiles

Profiles are named configuration bundles that tune every aspect of db0 for a specific workload. A chatbot needs fast decay and high recency. A coding assistant needs slow decay and high precision. Profiles let you configure this without touching individual parameters.

Built-in profiles

Profile	Use case	Scoring weights	Decay	Extraction
`PROFILE_CONVERSATIONAL`	Chat agents	0.5 sim / 0.35 rec / 0.15 pop	7 days	rules
`PROFILE_AGENT_CONTEXT`	Agent harnesses	0.65 / 0.2 / 0.15	14 days	rules
`PROFILE_KNOWLEDGE_BASE`	Document retrieval	0.8 / 0.1 / 0.1	30 days	rules
`PROFILE_CODING_ASSISTANT`	IDE and code tools	0.85 / 0.1 / 0.05	30 days	rules
`PROFILE_CURATED_MEMORY`	Human-authored files	0.9 / 0.05 / 0.05	90 days	manual
`PROFILE_HIGH_RECALL`	Benchmarks, research	0.6 / 0.25 / 0.15	14 days	rules
`PROFILE_MINIMAL`	Testing	similarity only	none	rules

Using a profile

import { db0, PROFILE_AGENT_CONTEXT } from "@db0-ai/core"

const harness = db0.harness({
  agentId: "my-agent",
  sessionId: "session-1",
  backend,
  profile: PROFILE_AGENT_CONTEXT,
})

Profile structure

Each profile configures five sections:

Ingest

Controls how raw content is processed into memories.

ingest: {
  mode: "chunk",              // "session" | "chunk" | "turn-context"
  chunkSize: 1600,            // max characters per chunk
  chunkOverlap: 320,          // overlap between chunks
  enrich: true,               // augment chunks with surrounding context
  enrichMode: "augment",      // "augment" | "rewrite"
  enrichWindowSize: 1,        // context window for enrichment
  latentBridging: false,      // create latent semantic bridges
}

Retrieval

Controls how memories are searched and scored.

retrieval: {
  topK: 8,                    // max results
  minScore: 0.4,              // minimum score threshold
  scoring: "hybrid",          // "similarity" | "hybrid" | "rrf"
  hybridWeights: {
    similarity: 0.65,
    recency: 0.2,
    popularity: 0.15,
  },
  decayHalfLifeDays: 14,     // recency decay half-life
  queryExpansion: false,       // expand query with related terms
  graphExpand: { maxExpand: 3 }, // traverse relationship edges
}

Extraction

Controls how facts are extracted from conversations.

extraction: {
  strategy: "rules",          // "rules" | "manual" | "llm"
  batchInterval: 5,           // buffer N turns before extracting
  maxBufferSize: 10,          // max buffer before force-extract
}

Context

Controls how memories are packed into the token window.

context: {
  budgetRatio: 0.6,           // fraction of token budget to use
  includeEdges: true,         // include relationship context
  maxPackItems: 20,           // max memories per pack
}

Reconciliation

Controls background maintenance.

reconciliation: {
  promotionThreshold: 5,      // access count to promote
  batchSize: 50,              // memories per reconcile batch
  autoReconcile: true,        // auto-run reconciliation
  reconcileInterval: 25,      // run every N turns
}

Custom profiles

You can create a custom profile by spreading a built-in one and overriding specific fields:

const myProfile = {
  ...PROFILE_AGENT_CONTEXT,
  retrieval: {
    ...PROFILE_AGENT_CONTEXT.retrieval,
    topK: 15,
    decayHalfLifeDays: 7,
  },
}