Core Concepts

DeltaMind has a small set of concepts. Understanding these makes everything else click.

Turn

A single message in a conversation. Has a turnId, role (user/assistant/system), and text content. Turns are the raw input. DeltaMind never modifies them.

Delta

A typed state change extracted from one or more turns. Not a summary — a structured mutation. DeltaMind defines 11 delta types:

Delta	Meaning
goal_set	What the session is trying to achieve
decision_made	A settled choice with rationale
decision_revised	A change to a prior decision
constraint_added	A rule, boundary, or requirement
constraint_revised	A relaxation, tightening, or amendment of a constraint
task_opened	Work to be done
task_closed	Work completed or abandoned
fact_learned	A stable piece of knowledge
hypothesis_introduced	A tentative idea — explicitly not a decision
branch_created	Unresolved alternatives (A vs B vs C)
item_superseded	Something replaced by something newer

Every delta carries a timestamp, source turn references, and a kind.

Candidate delta

The raw output of an extractor. A candidate is a delta that has not been reconciled yet. It includes:

• The proposed delta itself
• evidence — which turns and text snippets triggered it
• extractorConfidence — how confident the extractor is (0-1)
• semanticId — stable content hash for dedup

Candidates are proposals. The reconciler decides whether they become truth.

Active context state

The reconciled working set. A map of memory items plus a delta log and sequence counter. Each item has:

• id — unique within the session
• semanticId — stable content hash (same meaning = same hash)
• kind — one of: goal, decision, constraint, task, fact, hypothesis, open_question, rejected_option, dependency, risk, artifact
• summary — human-readable description
• status — active, tentative, superseded, resolved
• confidence — high, medium, low
• scope — session, project, durable
• sourceTurns — which turns produced this item (provenance)
• lastTouched — when the item was last created or modified
• tags — optional labels (e.g., “branch”)

Semantic ID

A stable content hash so that equivalent meaning converges regardless of which extractor found it or how it was phrased.

Algorithm:

1. Canonicalize the summary (lowercase, strip stop words, normalize verb forms)
2. Sort tokens alphabetically (order-independent)
3. FNV-1a 32-bit hash
4. Prefix with kind initial

Example: d-a3f7c012 — a decision with hash a3f7c012.

“Use TypeScript for type safety” and “using typescript for type safety” produce the same semantic ID. Word order doesn’t matter. Stop words are removed. Verb forms are normalized (using → use, building → build).

Provenance

The event log — what happened. Every accepted delta, every rejection, every checkpoint is recorded as a line in the provenance log. This is the flight recorder.

Three event types:

• Accepted — a delta that the reconciler accepted into state
• Rejected — a delta that was rejected (with reason)
• Checkpoint — a periodic snapshot of session health metrics

Snapshot

The current truth as machine-readable JSON. Versioned, deterministic (sorted keys), complete enough to resume a session. The snapshot is authoritative. Everything else is derived from it.

Projection

Human-readable markdown generated from state. Not authoritative. Never parsed back into state. Useful for inspection:

• ACTIVE_STATE.md — full working set overview
• DECISIONS.md — decisions ledger (active + superseded)
• TASKS.md — task tracker
• CONSTRAINTS.md — constraint ledger

Context export

A budgeted text rendering of the working set, designed for injection into LLM prompts. Priority ordering: constraints first, then decisions, goals, tasks, branches, recent deltas, changes since last export. Truncated at the character budget.

Advisory memory suggestion

DeltaMind can suggest updates to durable memory systems (like claude-memories files). These are advisory — the system proposes, a human or policy decides. Hypotheses and branch-tagged items are explicitly excluded from suggestions.