Patterns for sharing and isolating memory across multiple AI agents using ClawVault.

Multi-Agent Memory

When multiple agents operate in the same environment, memory management becomes a coordination problem. ClawVault supports both shared vaults (agents read and write the same memory) and isolated vaults (each agent has its own memory space).

Shared Vaults vs Isolated Vaults

Shared Vault

All agents read and write to the same vault directory. Best for agents that collaborate on the same tasks and need unified context.

Pros: Single source of truth, no sync overhead, full cross-agent visibility.

Cons: Risk of conflicting writes, noisy observations, harder to attribute memories.

Isolated Vaults

Each agent has its own vault. Cross-agent context is exchanged explicitly. Best for agents with distinct responsibilities.

Pros: Clean separation, no write conflicts, agent-specific tuning.

Cons: Requires explicit context sharing, potential for knowledge silos.

Share a vault when agents:

Work on the same project and need each other's decisions
Operate sequentially (one agent hands off to another)
Have non-overlapping categories (one stores decisions, another stores observations)

When to Isolate

Use separate vaults when agents:

Serve different users or contexts
Have conflicting observation schedules
Need different LLM configurations for compression
Operate in parallel on unrelated tasks

Agent-Specific Memory Folders

Within a shared vault, you can partition by agent using category prefixes or dedicated folders:

vault/
  .clawvault.json
  decisions/          # shared
  projects/           # shared
  agents/
    agent-alpha/      # agent-specific observations
    agent-beta/       # agent-specific observations
  inbox/
    agent-alpha/      # agent-specific captures
    agent-beta/

Each agent stores its observations and quick captures in its own subfolder while sharing access to decisions, projects, and other cross-cutting categories.

Configure per-agent paths in each agent's environment:

# Agent Alpha
export CLAWVAULT_AGENT_ID="agent-alpha"

# Agent Beta
export CLAWVAULT_AGENT_ID="agent-beta"

ClawVault uses CLAWVAULT_AGENT_ID to namespace observations and captures when set.

Cross-Agent Context

The clawvault context command generates task-relevant context from the vault. In a multi-agent setup, any agent can query the shared vault:

# Agent Beta needs context about a decision Agent Alpha made
clawvault context "authentication architecture decision"

This returns relevant memories regardless of which agent created them, because context searches across all vault categories.

Explicit Context Passing

For isolated vaults, use clawvault sync to share specific categories:

# Export decisions from Agent Alpha's vault to a shared location
clawvault sync /shared/decisions --categories decisions

# Agent Beta imports from the shared location
clawvault sync /shared/decisions --into /agent-beta/vault/shared-decisions

Patterns for Agent Orchestration

Hub-and-Spoke

A coordinator agent maintains the shared vault. Worker agents have isolated vaults but push summaries to the hub:

                ┌─────────────┐
                │ Coordinator │
                │   (hub)     │
                └──────┬──────┘
                       │ shared vault
            ┌──────────┼──────────┐
            │          │          │
      ┌─────┴──┐ ┌────┴───┐ ┌───┴─────┐
      │Worker A│ │Worker B│ │Worker C │
      │(isolat)│ │(isolat)│ │(isolat) │
      └────────┘ └────────┘ └─────────┘

Workers run clawvault sleep with summaries that the coordinator observes.

Pipeline

Agents operate sequentially. Each agent's sleep handoff becomes the next agent's wake input:

# Agent 1 finishes
clawvault sleep "research complete" --next "write draft based on findings"

# Agent 2 starts
clawvault wake  # picks up Agent 1's handoff

Parallel with Merge

Multiple agents work independently, then a merge agent consolidates:

# After parallel agents finish, merge agent runs:
clawvault observe --compress   # processes all new transcripts
clawvault link --all           # resolves cross-references
clawvault graph                # shows the unified picture

Multi-agent patterns are evolving. The CLAWVAULT_AGENT_ID environment variable and per-agent namespacing are the current recommended approach. Future versions may add first-class multi-agent coordination commands.

Multi-Agent Memory

Shared Vault

Isolated Vaults

On this page