Event Sourcing for AI Agent Memory: A Practical Guide

Event sourcing is the ideal storage model for AI agent memory because agents need the same thing auditors need: a complete, immutable history of everything that happened, queryable at any point in time. Here's how to wire it up with AllSource.

Why agents need event sourcing (not a vector database)

Most agent memory solutions store the current state — a vector embedding of the latest context. This is like keeping only the most recent database snapshot and throwing away the transaction log. You lose:

• Temporal reasoning: "What did I decide about this topic last week?"
• Provenance: "Why did I recommend that action?"
• Replay: "What would I have said if I'd known X at the time?"

Event sourcing stores every state change as an immutable event. The current state is derived — a projection over the event stream. You can always reconstruct any past state, replay decisions, and understand the full chain of reasoning.

The basic pattern: agent actions as events

Every meaningful agent action becomes an event:

# Agent observes something
curl -X POST https://api.all-source.xyz/api/v1/events \
  -H "Authorization: Bearer $API_KEY" \
  -d '{
    "event_type": "agent.observation",
    "entity_id": "agent-alpha",
    "payload": {
      "source": "user_conversation",
      "content": "User prefers concise answers with code examples",
      "confidence": 0.92,
      "session_id": "sess-789"
    }
  }'
 
# Agent makes a decision
curl -X POST https://api.all-source.xyz/api/v1/events \
  -d '{
    "event_type": "agent.decision",
    "entity_id": "agent-alpha",
    "payload": {
      "action": "switch_to_code_heavy_responses",
      "reasoning": "User preference detected: prefers code examples",
      "based_on": ["evt-001", "evt-003", "evt-007"]
    }
  }'

Each event captures what happened, why, and what evidence the decision was based on. This is the provenance chain that makes agent behavior auditable and debuggable.

Loading context: time-travel instead of RAG

Instead of embedding everything into a vector store and doing similarity search, you can load context by querying the event timeline:

# What does this agent know about the user's preferences?
curl "https://api.all-source.xyz/api/v1/events/query?\
entity_id=agent-alpha&\
event_type=agent.observation&\
limit=20&sort=desc" \
  -H "Authorization: Bearer $API_KEY"

For time-bounded context (e.g., "what happened in the last session"):

curl "https://api.all-source.xyz/api/v1/events/query?\
entity_id=agent-alpha&\
after=2026-04-20T09:00:00Z&\
before=2026-04-20T17:00:00Z" \
  -H "Authorization: Bearer $API_KEY"

This gives you deterministic, reproducible context loading — the same query always returns the same events. No embedding drift, no index staleness.

Projections: pre-computed agent state

For frequently accessed agent state (like user preferences), build a projection that stays in sync:

{
  "name": "user-preferences",
  "source": "agent.observation",
  "reduce": {
    "group_by": "entity_id",
    "keep_latest": true,
    "fields": ["content", "confidence"]
  }
}

The projection gives you 11.9μs lookups for the agent's current understanding of the user — no need to replay the full event stream on every request.

MCP integration: 43 tools for Claude

AllSource ships with 43 MCP tools that connect directly to Claude Desktop. Your agent can:

• query_events: Search the event stream by type, entity, or time range
• create_event: Store new observations, decisions, or actions
• list_projections: Check available pre-computed state
• get_snapshot: Load a point-in-time snapshot of any entity

This means Claude doesn't just store data in AllSource — it reasons about it using natural language queries over the event stream.

What this replaces

Getting started

# Get an API key (free tier: 100K events/month)
curl -X POST https://api.all-source.xyz/api/v1/onboard/start \
  -H "Content-Type: application/json" \
  -d '{"email":"you@example.com","name":"My Agent"}'

The response includes your API key and tenant ID. Start storing agent events immediately — no schema setup required (though we recommend it for production).

For the Prime module (knowledge graphs + vector search + compressed index), see the Prime platform page. For the core event store, start at all-source.xyz.