A working demonstration of shared context between humans and the agents they use.
Built on personal infrastructure.
An exhibition project by Mark Ferraz.
Shared memory wherever work happens
Captures land in a typed graph: people, decisions, tasks, knowledge, events, projects. Async workers classify each capture, infer relationships, and let stale relevance decay over time. Agents query the same graph over MCP. Humans and agents read from one source. That is the whole demonstration.
I built this for myself and the agents I work with. The same memory layer is exposed over MCP so other people and other agents can use it.
Connect Claude, ChatGPT MCP, Gemini, and internal agents to one memory plane. Query context, commit outcomes, and preserve continuity across every tool.
Main agents orchestrate strategy while subagents execute tightly scoped jobs with bounded tools and contextual slices.
Attach domain skills to agents, gate tools by responsibility, and keep capabilities explicit so delegation remains auditable.
Keep prompts lean by routing memory retrieval: fast paths for known patterns, deeper recall only when confidence or novelty requires it.
Every capture is typed and connected so agents retrieve usable context instead of raw logs. Scope follows project, role, and task intent.
Capture on mobile, monitor on web, and execute through MCP-enabled agents. Same memory graph, same context lineage, everywhere.
Control plane internals
LittleGuy is designed as a Memory Control Plane: structured memory, scoped execution, and portable context across web, mobile, and MCP clients.
Neo4j relationships + pgvector semantic recall in a dual-store architecture.
Decisions, tasks, people, events, documents, and more for higher-precision retrieval.
Token issuance, refresh, and revocation for secure agent-to-memory connectivity.
Deterministic, cached, and agentic retrieval modes keep token use and latency under control.
Parses raw capture into structured objects: entities, facts, commitments, and decisions.
Assigns node type and confidence so retrieval can route to relevant context classes.
Connects people, projects, and knowledge over time to maintain relationship-aware recall.
Reduces stale relevance unless reinforced, keeping memory fresh without manual cleanup.
Main agents delegate tasks to subagents with explicit boundaries for memory and tools, reducing accidental context bleed.
Skills and tool surfaces align to roles so each agent uses the right capabilities for the right class of work.
Recall policies prioritize high-signal context first, then expand only when required by novelty or ambiguity.
OAuth 2.0 + PKCE secures client connectivity while scoped tokens, revocation, and policy-aware tool surfaces keep memory access intentional. This is the foundation for reliable collaboration between humans, main agents, and delegated subagents.