Athena¶
The Memory Operating System for AI agents.¶
Perfect cross-session memory for every agent in your fleet. Athena captures every interaction, distills it into cognitive chains, and promotes what matters into a persistent knowledge graph. Your agents remember users, projects, and context across sessions, forever.
What Athena does¶
Every agent conversation evaporates when the session ends. Context windows overflow, users repeat themselves, and hard-won knowledge about a user's projects, tools, and struggles is lost. Athena eliminates that loss entirely.
Store every interaction through one API. Athena writes it to a short-term window your agent can read back instantly, then works in the background: a cognitive worker detects when the topic shifts, distills each topic into a summarized cognitive chain with its own embedding, and scores every chain with an Ebbinghaus-style heat model where memories that are recalled stay warm and memories that are ignored decay. Chains that stay hot are promoted into a long-term knowledge graph of identities, concepts, tools, and projects, connected by typed, confidence-weighted relationships. Ask Athena a question and it answers from all three tiers at once.
Athena is multi-tenant from the first byte. Every event is scoped to a tenant → user → agent hierarchy enforced at the auth layer, so one deployment serves your whole platform with hard data isolation.
STM
The hot window. A Redis-backed sliding window of recent events, dual-written to MongoDB for durability. Instant recall of the current conversation.
MTM
Cognitive chains. Topic-segmented summaries with embeddings in Milvus. Semantically searchable, heat-scored, merged and validated by a quality gate.
LTM
The knowledge graph. An ArangoDB graph of identities, concepts, tools, and projects. Typed relations, confidence weighting, community detection.
API & SDK
One surface, two protocols. gRPC and REST from a single proto definition, plus a Go client SDK. API-key, JWT, and mTLS auth built in.
Quick start¶
docker compose -f docker-compose.local.yml up -d
go run cmd/init-ltm/main.go
go run cmd/memory-server/main.go
REST listens on localhost:8080. gRPC listens on localhost:9090.
Where to start¶
Start with Architecture under the Concepts tab. It establishes the three-tier memory model, the background pipeline that moves memories between tiers, and the multi-tenant identity hierarchy. Every other page assumes that mental model.