Skip to main content

​
Memory

Bedrock agents have persistent memory that maintains context across runs. The memory system uses hierarchical summarization to keep long conversations within LLM token limits.

​
How Memory Works

Raw Messages
    ↓ (every 5 min)
5-Minute Summaries
    ↓ (every hour)
Hourly Summaries
    ↓ (every day)
Daily Summaries
    ↓ (every week)
Weekly Summaries
    ↓ (every month)
Monthly Summaries
As conversations grow, older details are compressed into higher-level summaries while recent context stays detailed.

​
Memory Context Structure

When an agent runs, it receives memory as two parts:

​
Historical Context (Cached)

Stable summaries from monthly → daily level. This content rarely changes and is cached by Claude for efficiency.

​
Recent Context (Dynamic)

Recent activity from hourly → 5-minute summaries plus raw messages. This changes frequently. 
┌─────────────────────────────────────┐
│         System Prompt               │ ← Cached
├─────────────────────────────────────┤
│     Historical Context              │ ← Cached
│  (Monthly/Weekly/Daily summaries)   │
├─────────────────────────────────────┤
│      Recent Context                 │ ← Not cached
│  (Hourly/5-min summaries + msgs)    │
├─────────────────────────────────────┤
│      Current Time + Wakeup          │ ← Not cached
└─────────────────────────────────────┘

​
Logging Messages

Log messages to agent memory: 
curl -X POST https://api.bedrock.orinlabs.org/api/memory/agents/AGENT_ID/log-messages/ \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "user", "content": "Hi, can you help me schedule a meeting?"},
      {"role": "assistant", "content": "Of course! When would you like to schedule it?"}
    ],
    "timestamp": "2024-01-15T10:30:00Z"
  }'
This is useful for:
  • Injecting external conversations (SMS, email) into memory
  • Seeding agents with initial context
  • Syncing state from external systems

​
Getting Memory Context

Retrieve the formatted memory context (what the agent sees): 
curl -X GET https://api.bedrock.orinlabs.org/api/memory/agents/AGENT_ID/memory-context/ \
  -H "Authorization: Bearer YOUR_API_KEY"
Response: 
{
  "context": "=== MONTHLY SUMMARY (December 2023) ===\nUser onboarded and set up task tracking...\n\n=== WEEKLY SUMMARY (Jan 8-14) ===\nDiscussed project deadlines...\n\n=== DAILY SUMMARY (Jan 15) ===\nUser asked about meeting scheduling..."
}

​
Getting Recent Messages

Retrieve raw recent messages (before summarization): 
curl -X GET https://api.bedrock.orinlabs.org/api/memory/agents/AGENT_ID/messages/ \
  -H "Authorization: Bearer YOUR_API_KEY"

​
Summarization Model

Memory summarization uses a separate, smaller model (default: claude-haiku-4-5) for efficiency. Configure per-agent: 
curl -X PATCH https://api.bedrock.orinlabs.org/api/cloud/agents/AGENT_ID/ \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "summarizer_model": "claude-haiku-4-5"
  }'

​
Resetting Memory

Clear all memory for an agent: 
curl -X POST https://api.bedrock.orinlabs.org/api/cloud/agents/AGENT_ID/reset-memory/ \
  -H "Authorization: Bearer YOUR_API_KEY"
This permanently deletes all messages and summaries. The agent will start fresh with no memory of past interactions.

​
Memory and Tool Calls

When agents make tool calls, the calls and results are automatically logged to memory: 
[
  {
    "type": "function_call",
    "name": "list_tasks",
    "arguments": "{\"limit\": 10}"
  },
  {
    "type": "function_call_output",
    "output": "Found 3 tasks:\n1. Review proposal (due tomorrow)\n2. Call client..."
  }
]
This means agents remember what tools they’ve used and what they learned.

​
Memory Timestamp Handling

Messages can include explicit timestamps: 
{
  "messages": [...],
  "timestamp": "2024-01-15T10:30:00Z"
}
If omitted, the current time is used. Timestamps affect which summary bucket messages fall into.

​
Prompt Caching Benefits

The hierarchical memory structure enables prompt caching with Claude:
Content TypeCache StatusWhy
Tool descriptionsCachedRarely changes
System promptCachedStatic
Historical memoryCachedChanges slowly
Recent memoryNot cachedChanges every run
This can reduce costs by 50-80% for long-running agents.

​
Best Practices

Let It Summarize

Don’t manually manage memory - let the summarization system work.

Use Timestamps

Include accurate timestamps when logging external messages.

Reset Sparingly

Only reset memory when truly needed - context is valuable.

Check Context

Use memory-context endpoint to debug what agents see.