A place of
memory
for your agents.
Gradatum is a self-hosted, embedded memory backbone for multi-agent AI systems. Today: Vault — sovereign memory, queryable via MCP stub. Tomorrow: identity, coordination, compute.
Modules
Memory for your agents
Vault
Embedded memory backbone. Cures memory rot in LLMs and lets Claude, Gemini, Codex and your home-grown agents share what they know — across sessions, across machines, across runs.
- SQLite FTS5 (BM25) + bge-m3 1024d embeddings
- BM25 + cosine + reranker (no-op by default, ONNX optional) — RRF (Reciprocal Rank Fusion)
- Multi-signal scoring (recency × pagerank backlinks)
- MCP stdio + REST :19090 + auto-refresh JWT
Agents Context layer
Gateway + Engine · Context runtime
The layer that orchestrates memory and compute for your agents. Two of three foundations are already in place — the runtime that ties them together is the next milestone.
- Gateway — unified LLM routing by alias, circuit-breaker
- Engine fleet — local inference, no SaaS round-trip
- Context assembly — sliding window, proactive recall
- ReAct runtime — agent loop + guards
One discipline.
Many primitives.
Modern agent stacks are stitched from a dozen vendor SaaS, each with its own auth, its own latency, its own privacy story. Gradatum builds the opposite — starting with sovereign memory: self-hosted, embedded, MCP-native, owned end-to-end. Identity, coordination, and compute follow on the roadmap.
cargo install · systemctl enable · your data stays here