Agentic memory
Memory is crowded. Agentic coding is solved. Agentic memory doesn’t exist.
Every AI memory tool on the market is storage. A pile of text chunks with a vector index on top. That is not memory any more than a filing cabinet is a brain.
Agentic memory is different. It is context that agents can build against the way an engineer builds against a database — scoped, typed, persistent, queryable. Structure rigid enough that a workflow can depend on it, flexible enough to evolve with the work.
That does not exist in the market. Until Brian.
What Brian is
Brian is a typed memory graph for Claude, delivered over the Model Context Protocol.
- Nodes, edges, blobs, models. A four-table semantic architecture, not a chunk store.
- Spaces.Isolated memory domains — work, health, finance, clients — retrieved only when relevant.
- Sessions. Tracked work with a full audit trail of what was decided and why.
- Souls. Persistent behaviour definitions that make agents reliable across runs.
- Skills. Procedural knowledge that teaches Claude how to use the graph.
Connect once via MCP. Every Claude conversation from that point starts with the context that matters and ignores the rest.
What changes when memory is agentic
When context is scoped and persistent, the stack collapses. You stop re-explaining. You stop pasting yesterday’s decisions into today’s prompt. You stop watching Claude lose the plot halfway through a long project.
In a single afternoon this week I architected a bloodwork analysis product end-to-end with Claude, pulling from memory that already held the business context, the technical stack, the prior reasoning. Not because Claude got faster. Because the context held.
Brian + Claude is the first stack where context compounds instead of resets.
Proof
- Hybrid retrieval over vector and keyword. RAGAS faithfulness 0.92. Zero hallucinations on the golden test set.
- Semantic chunking with heading-based structure and passage-level retrieval.
- Typed document pipeline that ingests, chunks, and indexes Google Drive contents.
- Agent framework with typed souls, session audit trail, and inter-agent communication.
- Hybrid retrieval, not just vectors. Typed nodes, not just chunks. Tracked sessions, not just stored chats.
How it compares
| Storage tools | Brian | |
|---|---|---|
| Retrieval | Vector similarity | Hybrid: vector + keyword + typed graph |
| Structure | Flat chunks | Typed nodes, edges, blobs, models |
| Context scope | Global blob | Spaces, sessions, souls |
| Agent behaviour | Stateless per run | Persistent via souls |
| Audit trail | None | Every session tracked |
Mem0, Zep, MemPalace, native AI memory — all storage. Brian is memory.
Start
Brian is live at mcp.imbrian.io. Free tier available. Pro at $15 per month.
Five minutes to a Claude that remembers everything.