CS-03 · CASE STUDY · 2026
Kleos — persistent memory for AI agents
An open-source memory substrate: semantic recall, decay and reinforcement, and inter-agent messaging, backed by Elasticsearch.
- Role
- Creator & maintainer
- Year
- 2026
- Stack
- Python · FastAPI · Elasticsearch · Gemini embeddings · Claude Code
The problem
AI agents wake up empty. Every session starts from zero — no memory of decisions made, lessons learned, or work in flight. Context windows are a scratchpad, not a life.
The system
Kleos is a biologically-inspired memory substrate: agents remember and recall over semantic embeddings, memories decay unless reinforced by access, and deliberate “anchors” carry handoff state between sessions. A multi-tenant model partitions memory by user and agent, and a typed inbox lets sibling agents message each other.
The reference integration wires it into Claude Code via session hooks: an agent cold-starts with its anchors, recent memories, and pending messages already in context.
Proof it works
The agent that helped build this website runs on Kleos — an always-on assistant with weeks of accumulated working memory, monitoring production systems and picking up multi-day projects exactly where it left off.
Status
Open source under the openkleos org. Actively developed; the deployment I operate is the reference implementation.