Honcho is an open-source memory layer for building stateful AI agents, created by Plastic Labs. Instead of treating memory as a pile of stored facts, Honcho continuously reasons over conversation history to build conclusions about each user, so an agent can start every session already knowing what matters. The project has about 7,700 stars on GitHub, is released under the AGPL-3.0 license, and was actively developed in 2026.
You write messages to Honcho, and two things happen: they are stored and indexed, and a reasoning engine (called Neuromancer) learns from them in the background, drawing and weighting conclusions about the people and agents involved โ the “peers” in a session. When your application needs context, a single context() call returns curated reasoning plus relevant history within a token budget you set, which the team says cuts prompt sizes dramatically compared with dumping raw transcripts. Background “dreaming” processes keep refining that understanding between sessions, and a chat() endpoint answers specific questions about a peer on demand. Honcho is model-agnostic and can be self-hosted or used through the managed API.
Freemium. The Honcho server is open source (AGPL-3.0) and can be self-hosted for free. The managed Honcho API charges per use โ ingestion with reasoning is priced per million tokens and deeper reasoning queries are billed per query โ with startup credits and custom enterprise plans available.
Agent builders whose products live or die by continuity: AI companions, coding assistants, tutors, support agents and games where characters should remember relationships and preferences across sessions.