Skip to content
DimpleDimple

Architecture

Dimple is a local-first memory system: everything is a memory unit (text, code, facts — one uniform artifact), organized by a topic tree (where to search), connected by an artifact graph (what is near, what explains, what contradicts or supersedes), annotated by concepts (emergent subject tags), and maintained by a durable job runtime (all structural work, idempotent and resume-safe).

                     ┌────────────────────────────────┐
                     │             dimple             │
                     │  write ──▶ embed ──▶ tree      │
                     │  query ──▶ descend ──▶ hybrid   │
                     └───────┬─────────────┬──────────┘
                             │             │
                    ┌────────▼─────┐  ┌────▼─────────┐
                    │  dimple.db   │  │   jobs.db    │
                    │ memories     │  │ durable queue│
                    │ topic tree   │  │ splits/merge │
                    │ graph + audit│  │ dreams/repair│
                    └──────────────┘  └──────────────┘

Reading map (dependency order)

Section titled “Reading map (dependency order)”
Page Answers
Memory units & lifecycle What is stored, what states it can be in, what “superseded” means
Topic tree How memories are organized so search cost grows with depth, not count — routing, splits, merges, rebalance
Retrieval How queries turn into ranked, scoped, audited results — hybrid fusion, beam descent, budgets
Artifact graph How memories relate — supports / supersedes / contradicts / code_calls, the pending→confirm lifecycle
Concepts & consolidation How the corpus organizes itself — deterministic clustering + annotations + query sovereignty
Dream (consolidation) The LLM pass — edge classification, concept clustering, provenance, gates
Durability & determinism What keeps everything consistent across restarts — the job runtime and the reproducibility contract

Every page opens with “In one sentence” and closes with “Key facts” (stable invariants + hard limits) so both people and agents can extract the contract quickly. Algorithms are explained where they are used, not in an appendix.

Storage at a glance

Section titled “Storage at a glance”

Two SQLite files (or one Turso remote database — the same engine over the wire):

Database Holds
dimple.db (storeUrl) memories, topic tree, graph edges, concepts, retrieval audit
jobs.db (jobs.dbUrl) the durable job queue + workflow state

Algorithms at a glance

Section titled “Algorithms at a glance”
Mechanism Where See
Greedy centroid descent (stable tie-breaks) insert routing Topic tree → Insert routing
CF-triplet sufficient statistics centroids/radii (splits, rebalance, clustering) Topic tree → Splits
Farthest-pair seeds → Lloyd refinement leaf/internal splits Topic tree → Splits
Donor-min borrow, nearest-sibling merge underflow Topic tree → Merge & borrow
RRF fusion (k = 60) hybrid retrieval Retrieval → Hybrid fusion
Beam descent tree-scoped queries Retrieval → Tree descent
FNV-1a query hashing audit identity Retrieval → Audit
Deterministic k-means (corpus pass) concept clustering Concepts → Clustering
LLM edge classification + provenance rule consolidation Dream → What it does