What DeltaMind Is
DeltaMind is active context compaction for long-running AI sessions. It replaces transcript-as-memory with state-as-memory.
The short version
Conversations produce state changes: decisions get made, constraints get added, tasks open and close, hypotheses emerge and collapse. DeltaMind tracks those changes as typed deltas and reconciles them into a structured, queryable working set.
The result: a 500-turn session feels clearer at turn 500 than at turn 50.
What it does
- Emits typed deltas from conversation turns (decisions, constraints, tasks, goals, hypotheses, revisions)
- Reconciles deltas into a structured state with provenance tracking
- Exports budgeted context for injection into LLM prompts
- Persists state as snapshots for session resume
- Provides operator tools for inspection and debugging
What it is not
- Not a chat UI. DeltaMind is a library and CLI. It processes turns that come from somewhere else.
- Not a vector database. There is no embedding, no similarity search, no retrieval augmentation. State is structured, not vectorized.
- Not a generic note-taking app. It tracks conversation state changes, not arbitrary documents.
- Not a durable memory system by itself. Session state is authoritative within a session. Durable memory promotion is advisory — the system suggests updates, it does not autonomously write them.
- Not a replacement for reasoning models. DeltaMind manages what the model remembers, not how it thinks.
The thesis
Transcripts grow linearly. Every turn adds text regardless of whether anything meaningful changed. But state changes are sparse — most turns are elaboration, not mutation.
DeltaMind exploits this sparsity. On long sessions (56-62 turns), context is 12-24% of raw transcript size. The longer the session, the better the compression ratio.
Summaries don’t solve this. They flatten nuance, destroy provenance, and merge speculation with settled truth. You can’t ask a summary “what did we decide about X and why?” You can ask DeltaMind exactly that.
Packages
| Package | Purpose |
|---|---|
@deltamind/core | Typed deltas, state model, extraction, reconciliation, persistence, adapters |
@deltamind/cli | Operator CLI for session inspection, export, replay, and debugging |
Where to start
- Core Concepts — the vocabulary you need
- Session Runtime — how to use DeltaMind in code
- Four-Pass Engine — how extraction works under the hood
- CLI Reference — operator commands