rlm-memory

name	rlm-memory
description	Long-term memory persistence for RLM using Daytona durable volume storage. Use when storing, recalling, listing, or searching data that persists across sandbox sessions in the daytona_pilot runtime.

⚠️ CRITICAL: /tmp/ is ephemeral — data is lost when the sandbox restarts. ALWAYS write persistent data to /home/daytona/memory/ using a volume_name. ALWAYS call interp.shutdown() in a finally block.

RLM Memory — Persistent Storage

Persist data across sandbox sessions using the Daytona mounted durable volume.

There are no slash commands — all interactions use the Python API.

---

Daytona Durable Volume

The Daytona mounted volume is rooted at /home/daytona/memory/ inside the sandbox.

Canonical Directories

Directory	Purpose
/home/daytona/memory/memory/	Key-value and named memory items
/home/daytona/memory/artifacts/	Produced outputs and saved results
/home/daytona/memory/buffers/	Named buffer lists (session logs, staging)
/home/daytona/memory/meta/	Session manifests and workspace metadata

Session Manifest Path

/home/daytona/memory/meta/workspaces/<workspace_id>/users/<user_id>/react-session-<session_id>.json

---

Store and Recall

from fleet_rlm.integrations.providers.daytona.interpreter import DaytonaInterpreter

interp = DaytonaInterpreter(
    repo_url="https://github.com/your-org/your-repo",
    timeout=120,
    volume_name="rlm-volume-dspy",
)
interp.start()
try:
    interp.execute('''
import json, os
os.makedirs('/home/daytona/memory/artifacts', exist_ok=True)
data = {'result': 'my-finding', 'created': '2026-04-03'}
with open('/home/daytona/memory/artifacts/analysis.json', 'w') as f:
    json.dump(data, f)
''')

    result = interp.execute('''
import json
with open('/home/daytona/memory/artifacts/analysis.json') as f:
    data = json.load(f)
SUBMIT(result=data['result'])
''')
    print(result.result)
finally:
    interp.shutdown()

---

Buffer Pattern

add_buffer(name, value) / get_buffer(name) are injected by the sandbox driver (runtime/execution/sandbox_assets.py) and persist named lists across execute() calls within the same sandbox session:

interp.start()
try:
    interp.execute('add_buffer("log", "Step 1: setup")')
    interp.execute('add_buffer("log", "Step 2: done")')
    result = interp.execute('SUBMIT(log=get_buffer("log"))')
    print(result.log)  # ["Step 1: setup", "Step 2: done"]
finally:
    interp.shutdown()

---

Volume-Aware Document Tools (Inside Sandbox)

Available inside sandbox code on the Daytona path:

• load_text_from_volume(path) — loads durable mounted-volume content
• load_from_volume(path) — loads file contents from volume
• process_document(...) — targeted at mounted-volume content
• workspace_read(path) — low-level transient workspace helper (not durable)

# Use load_text_from_volume or load_from_volume for durable volume reads:
result = interp.execute("""
content = load_text_from_volume('/home/daytona/memory/artifacts/report.txt')
SUBMIT(size=len(content))
""")
print(result.size)

---

Checkpoint Pattern

interp.start()
try:
    # Save intermediate results during long workflows
    interp.execute('''
import json, os
root = '/home/daytona/memory/buffers/checkpoints'
os.makedirs(root, exist_ok=True)
with open(f'{root}/batch_1.json', 'w') as f:
    json.dump([{'i': i} for i in range(100)], f)
''')

    # Resume in next execute() call or next session
    result = interp.execute('''
import json
with open('/home/daytona/memory/buffers/checkpoints/batch_1.json') as f:
    previous = json.load(f)
SUBMIT(count=len(previous))
''')
    print(result.count)  # 100
finally:
    interp.shutdown()

---

Troubleshooting

See rlm-debug for comprehensive diagnostics and daytona-runtime for Daytona volume specifics.

Common Mistakes

• Never use /tmp/ for data that must persist — it is cleared on sandbox restart
• Always use try/finally to ensure interp.shutdown() is called
• Never use different volume_name values across sessions that must share data
• Do not use pickle for structured data — use json.dump / json.load