Language models are stateless. Most products called memory compensate by embedding text and returning nearby chunks. That is useful retrieval, but it does not answer the harder questions: what was true then, what is true now, which source is authoritative, what changed, who may see it, and why a memory was selected.
LongMemory treats those questions as the memory system itself.
- Immutable content with recorded-time and valid-time history.
- Executable typed edges, entities, worlds, grounding, contradiction, and provenance.
- Strict, historical, associative, world-grounded, and multilingual recall.
- Deterministic decay and explicit reinforcement without rewriting source truth.
- Project-scoped decisions, tasks, Skills, Chat Memory, LLM-Wiki, and CodeGraph assets.
- Explainable evidence selection under token budgets and access policy.
- One local-first engine across npm, CLI, HTTP, MCP, dashboard, and editor workflows.
It is not a hosted black box, a transcript dump, an autonomous permission system, or a replacement for model context. It is a governed evidence layer that lets applications and agents carry durable state without surrendering ownership or auditability.
The model can remain stateless. The application no longer has to be amnesiac.