Every interaction stays source-backed.
Conversations, tool calls, documents, and agent actions land as durable memory events with timestamps, scope, provenance, and metadata.
MongoDB-native agent memory
Memorythat keepsits evidence.
Memongo gives AI agents durable long-term memory: event history, structured facts, knowledge chunks, procedures, graph context, and hybrid retrieval in one inspectable system.
eventsvectorgraphtext
one memory storesource, search, graph, proof
Storesevents, facts, procedures, docs
Retrievesvector, lexical, hybrid, graph
Explainssources, scores, roles, timestamps
The hidden hard part
Memory is not a vector table.
A useful agent needs more than nearest neighbors. It needs the actual event, the current fact, the old fact it replaced, the exact name a user typed, the relationship between sessions, and the proof that a context bundle was assembled honestly.
Facts become current state, not loose notes.
Preferences, procedures, profile details, revisions, and superseded values are modeled so an agent can know what changed and what still holds.
Semantic, lexical, graph, and hybrid recall work together.
Vector similarity finds meaning, full-text search catches names and exact facts, graph links recover relationships, and hybrid ranking chooses the right evidence.
The answer can point back to the memory.
Context bundles preserve source IDs, scores, roles, timestamps, and stale/current labels so memory can be inspected instead of trusted blindly.
The shape of the system
One memory substrate, many recall modes.
Memongo is built around a simple belief: agent memory should live where documents, indexes, relationships, operational queries, and provenance can be reasoned about together.
Document memory
Events, facts, KB chunks, procedures, graph edges, and telemetry live together.
Vector recall
High-recall semantic search for fuzzy questions and long-running context.
Lexical recall
Exact names, dates, identifiers, and proper nouns stay recoverable.
Hybrid ranking
Semantic and keyword evidence can be fused without leaving the memory store.
Graph context
Episodes, entities, sessions, and scopes become traversable relationships.
Operational truth
Health, indexes, provenance, cleanup, and release gates are first-class.
Not benchmark theater
Built for audit before bragging.
Memongo keeps benchmark claims scoped. Retrieval evidence and judged answer quality are separated. Source IDs, commands, metadata, topology, cleanup proof, and model posture matter more than a headline.
Public postureSelected retrieval evidence is published. Broad ecosystem leadership is not claimed.
The product is open source now. The benchmark work remains honest, reproducible, and deliberately scoped.
Use it like infrastructure
Add memory, search memory, inspect the answer.
Run the API, connect the SDK, then let your agent retrieve context from the same place that stores the source evidence.
const memongo = new MemongoClient({
baseUrl: "http://127.0.0.1:3847"
})
await memongo.add({
sessionId: "agent-42",
content: "Romi prefers concise release notes."
})
const context = await memongo.search({
sessionKey: "agent-42",
query: "How should I write the launch note?",
limit: 8
})