Multi-Agent Memory Sharing Patterns Explained

Pattern 1: Shared Memory Bus

The shared bus is the simplest pattern: all agents read from and write to a single memory store with no access controls or namespacing. Agent A stores an observation, and Agent B can immediately retrieve it.

This pattern works well when agents have complementary, non-overlapping roles and all produce high-quality observations. A monitoring agent and an execution agent working on the same system benefit from full sharing because their observations are complementary: the monitoring agent discovers the problem, the execution agent fixes it. Neither agent needs protection from the other's output.

The shared bus breaks down when agents produce observations of varying quality, when many agents write frequently, or when agents work on different topics that should not cross-pollinate. A research agent generating speculative hypotheses and a compliance agent recording regulatory requirements should not share a flat namespace because the speculative hypotheses would appear alongside regulatory requirements in retrieval results, making it harder for any agent to find what it needs.

Best for: Small agent teams (2 to 4 agents) with complementary roles and consistently high-quality output.

Pattern 2: Scoped Namespaces

Scoped namespaces give each agent its own write space while allowing configurable read access to other agents' namespaces. Agent A writes to namespace "research," Agent B writes to namespace "execution," and both can read from both namespaces plus a "shared" namespace that contains promoted, high-confidence findings.

The key advantage is write isolation. A verbose research agent that stores 50 speculative observations does not pollute the execution agent's retrieval when the execution agent queries its own namespace. The execution agent can optionally include the research namespace in its queries when it needs broader context, but this is explicit rather than default.

Namespace design follows agent roles. Each role gets a namespace: "monitoring," "research," "execution," "planning," "compliance." A "shared" or "validated" namespace contains findings that have been promoted from individual namespaces after reaching a confidence threshold. Promotion can be automatic (confidence above 0.85), manual (a coordinator agent reviews and promotes), or collaborative (the same finding appears in two or more agent namespaces, indicating independent corroboration).

Best for: Medium-sized agent teams (4 to 10 agents) with defined roles and varying output quality.

Pattern 3: Blackboard Architecture

The blackboard pattern, originally from classical AI expert systems, structures shared memory as a workspace where agents post observations (write) and read the current state of the workspace to decide what to do next. A control component monitors the blackboard and activates agents based on the current state. When the monitoring agent posts "latency spike detected," the control component activates the diagnostic agent. When the diagnostic agent posts its findings, the control component activates the remediation agent.

The blackboard differs from a shared bus in two important ways. First, it has structure: the workspace is organized into sections (hypothesis, evidence, actions, results) rather than a flat list. Second, it has a controller: a coordinating process that determines which agent acts next based on what is currently on the blackboard. This makes the blackboard suitable for collaborative problem-solving where the order of agent execution matters and different agents contribute different expertise.

The trade-off is complexity. The blackboard architecture requires a coordination layer that monitors the workspace, decides which agent to activate, and manages the flow of execution. This coordination layer is itself a component that can fail, and debugging issues requires understanding both the agent behavior and the coordination logic. For simple agent teams where agents can run independently, the blackboard adds overhead without proportional benefit.

Best for: Complex collaborative problem-solving where agent execution order matters and agents contribute different expertise to a shared solution.

Pattern 4: Hierarchical Memory

Hierarchical memory organizes access into tiers: individual (only this agent), team (agents in the same role group), and organization (all agents). Each tier has different access patterns and different quality thresholds.

The individual tier is the agent's scratch space. It stores speculative observations, intermediate results, and working notes. Quality is not filtered because only the originating agent accesses this tier. The team tier is shared among agents with the same role. A group of monitoring agents shares their individual findings so they can benefit from each other's coverage. Promotion to the team tier requires a minimum confidence threshold. The organization tier is the global shared memory that all agents can access. Promotion to this tier requires high confidence, corroboration from multiple sources, or explicit approval from a coordinator.

This pattern maps to how knowledge flows in human organizations. An individual researcher has rough notes (individual tier). The research team has a shared findings document (team tier). The company has a wiki that contains validated, reviewed knowledge (organization tier). Each tier has different signal-to-noise ratios and different access patterns.

Best for: Large agent systems (10+ agents) with multiple role groups and significant differences in observation quality.

Choosing the Right Pattern

Start with the simplest pattern that meets your needs. Most agent systems start as 2 to 3 agents with complementary roles, which works fine with a shared bus. Add namespacing when retrieval quality degrades because of noisy or irrelevant results from other agents. Move to a blackboard when agent execution order matters for correctness. Move to hierarchical memory when you have multiple agent groups that need different levels of access.

Adaptive Recall supports all four patterns through its flexible tagging and filtering system. Agents tag memories with their ID, role, and confidence. Recall queries filter by any combination of tags, implementing namespace-like isolation. The knowledge graph provides blackboard-like coordination by connecting entities across agent observations. And the confidence-based recall threshold naturally implements hierarchical promotion, since low-confidence observations only appear in unfiltered queries while high-confidence findings surface in any query.