Open source

Store what changed.

Active context compaction for AI agents. Typed deltas replace transcript sprawl with structured, queryable state that stays compact as conversations grow.

Get started Read the Handbook

Session

const session = createSession(); session.ingestBatch(turns); await session.process();

Query

session.stats(); // { items: 7, decisions: 2, // constraints: 3, openTasks: 1 }

Export

session.exportContext({ maxChars: 2000 }); // Budgeted working set for injection

Why DeltaMind

Transcripts grow linearly. State grows sublinearly.

Typed Deltas

11 delta types capture decisions, constraints, tasks, goals, hypotheses, and revisions. No freeform summaries that smear meaning.

Hybrid Extraction

Rule-based precision meets LLM semantic recall. Both compute stable semantic IDs — equivalent meaning converges regardless of path.

Safety First

Zero canonization of speculation. Rejections over corruption. Advisory-only memory promotion. The system would rather abstain than poison the ledger.

Architecture

Four-pass pipeline, three representations, single source of truth.

Event Gate

Cheap heuristics filter turns worth analyzing. High recall, tolerant of false positives. Saves extraction cost on chatter.

Extract + Reconcile

Extractors emit candidates. The reconciler decides truth. Duplicates are merged, superseded items are marked, provenance is preserved.

Persistence

Event log (what happened), snapshot (current truth), markdown (human inspection). Three formats, each with one job.

Scaling Results

The longer the session, the more DeltaMind earns its keep.

Metric

Short (9-14 turns)

Long (56-62 turns)

Context vs raw

18-62%

12-24%

Item growth rate

Linear

Sublinear (2.9x for 5x turns)

Query score

6/6

Provenance

Full

Usage

Install

npm install @deltamind/core @deltamind/cli

Create a session

import { createSession } from "@deltamind/core";

const session = createSession({ forceRuleOnly: true });

session.ingestBatch([
  { turnId: "t-1", role: "user",
    content: "Build a REST API with TypeScript." },
  { turnId: "t-2", role: "assistant",
    content: "Setting up Express with TypeScript." },
  { turnId: "t-3", role: "user",
    content: "Switch to Fastify instead." },
]);

await session.process();

Query and export

// What's in state?
const stats = session.stats();
// { totalItems: 5, activeDecisions: 1, ... }

// Budgeted context for injection
const ctx = session.exportContext({ maxChars: 2000 });

// Save for later
const snapshot = session.save();

CLI operations

deltamind inspect              # State by kind
deltamind changed --since 5    # What changed
deltamind explain item-3       # Deep-dive
deltamind export --for ai-loadout
deltamind suggest-memory       # Advisory updates
deltamind save && deltamind resume

Delta Types

11 typed state changes — not summaries.

Decisions & Goals

goal_set, decision_made, decision_revised. Settled choices with provenance. Revisions supersede prior decisions while preserving history.

Constraints & Tasks

constraint_added, constraint_revised (relaxed/tightened/amended), task_opened, task_closed. Boundaries and work items with lifecycle tracking.

Knowledge & Uncertainty

fact_learned, hypothesis_introduced, branch_created, item_superseded. Facts are stable. Hypotheses stay tentative until explicitly promoted.