Give Your Claude Agent a Memory That Survives Sessions: The Managed Agents Memory Store

Chris Harper

3 min read

Jul 22, 2026 · 04:05 UTC

Workflow

Agents

Claude Code

Best Practices

TL;DR: Attach a memory store to a Claude Managed Agent session and the agent gets a persistent /mnt/memory/ directory it reads and writes with ordinary bash tools — every write versioned, surviving across sessions.

Every agent conversation ends and context clears. The checkpointer pattern (thread IDs) gives you replay within a thread, but not recall across threads or weeks. Claude Managed Agents ships a simpler answer: a memory store is a directory the agent reads and writes with the same bash and file tools it already uses, persisted and versioned server-side. Today's SDK releases (Python 0.116.0, TypeScript 0.110.0, Go 1.56.0, Java 2.48.0, Ruby 1.55.0, PHP 0.36.0, C# 12.35.0) ship the new agent-memory-2026-07-22 beta header and wire it in automatically on all memory store calls.

Create and mount a memory store

import anthropic

client = anthropic.Anthropic()

# 1. Create a workspace-scoped memory store (one-time setup)
store = client.beta.memory_stores.create(name="user-preferences")
store_id = store.id   # e.g. "memstore_01AbC..."

# 2. Mount it in a session — the agent sees it at /mnt/memory/
session = client.beta.agents.sessions.create(
    model="claude-opus-4-8",
    system=(
        "You are a personal assistant. Your long-term memory is at "
        "/mnt/memory/. Check it at the start of each conversation and "
        "update it when you learn something new about the user."
    ),
    resources=[
        {
            "type": "memory_store",
            "memory_store_id": store_id,
            "mount": "/mnt/memory",
            "access": "read_write",   # or "read_only" for consumer agents
        }
    ],
    tools=[{"type": "agent_toolset_20260401"}],
)

What the agent does with it

The agent uses its existing file tools — no new API to learn:

# Agent writes at the end of a session
cat > /mnt/memory/style.md << 'EOF'
User: prefers bullet points, dislikes long explanations, timezone UTC-6
EOF

# Agent reads at the start of the next session (different thread)
cat /mnt/memory/style.md

The same store mounts in the next session (new thread, days later). The agent opens /mnt/memory/style.md and already knows.

Every write is versioned

# Every mutation creates an immutable memver_... ID
versions = client.beta.memory_stores.versions.list(store_id)
for v in versions.data:
    print(f"{v.id} — {v.created_at} — session {v.session_id}")

# Roll back a bad write
client.beta.memory_stores.restore(store_id, version_id="memver_01XyZ...")

Multi-agent memory sharing

Mount the same store_id in two sessions with read_only on the consumer to share what one agent learned:

resources=[
    {"type": "memory_store", "memory_store_id": store_id, "mount": "/mnt/memory", "access": "read_only"},
]

A research agent writes; a summarizer agent reads — no external message bus needed.

Beta header note

Memory store endpoints use agent-memory-2026-07-22 instead of managed-agents-2026-04-01. The updated SDKs set the correct header automatically. If you set beta headers manually, replace managed-agents-2026-04-01 with agent-memory-2026-07-22 on memory store calls — sending both returns a 400.