feat: add memory backend

.github/dependabot.yml

@@ -30,6 +30,7 @@ updates:
       - "/examples/deep_research"
       - "/examples/nvidia_deep_agent"
       - "/examples/text-to-sql-agent"
+      - "/examples/memory-backend-agent"
     schedule:
       interval: "monthly"
     groups:

README.md

@@ -27,11 +27,16 @@ Deep Agents is an agent harness. An opinionated, ready-to-run agent out of the b
 
 - **Planning** — `write_todos` for task breakdown and progress tracking
 - **Filesystem** — `read_file`, `write_file`, `edit_file`, `ls`, `glob`, `grep` for reading and writing context
+- **Path-keyed virtual files (`MemoryBackend`)** — The same file tools can target **path-keyed records** instead of local disk. Implement `PathMemoryStore`, or route a prefix through `CompositeBackend` (for example `/memories/` → memory, `/` → filesystem).
+- **PowerMem (reference store)** — `PowerMemPathStore` wires PowerMem into `MemoryBackend` for durable, multi-tenant storage without changing tool names. See [`examples/memory-backend-agent/`](examples/memory-backend-agent/) for persistence and optional semantic search tools.
 - **Shell access** — `execute` for running commands (with sandboxing)
 - **Sub-agents** — `task` for delegating work with isolated context windows
 - **Smart defaults** — Prompts that teach the model how to use these tools effectively
 - **Context management** — Auto-summarization when conversations get long, large outputs saved to files
 
+> [!NOTE]
+> **Two kinds of "memory"** — Loading long-lived instructions from files such as `AGENTS.md` into the system prompt is separate from **path-keyed virtual files** (`MemoryBackend`). You can use both together.
+
 > [!NOTE]
 > Looking for the JS/TS library? Check out [deepagents.js](https://github.com/langchain-ai/deepagentsjs).
 
@@ -71,6 +76,8 @@ agent = create_deep_agent(
 
 MCP is supported via [`langchain-mcp-adapters`](https://github.com/langchain-ai/langchain-mcp-adapters).
 
+For **path-keyed virtual files** with an in-memory store, optional **PowerMem** persistence, and extra tools for semantic recall, see [`examples/memory-backend-agent/`](examples/memory-backend-agent/).
+
 ## Deep Agents CLI
 
 A pre-built coding agent in your terminal — similar to Claude Code or Cursor — powered by any LLM. One install command and you're up and running.
@@ -103,7 +110,7 @@ See the [CLI documentation](https://docs.langchain.com/oss/python/deepagents/cli
 - **100% open source** — MIT licensed, fully extensible
 - **Provider agnostic** — Works with any Large Language Model that supports tool calling, including both frontier and open models
 - **Built on LangGraph** — Production-ready runtime with streaming, persistence, and checkpointing
-- **Batteries included** — Planning, file access, sub-agents, and context management work out of the box
+- **Batteries included** — Planning, file access, sub-agents, and context management work out of the box; add `MemoryBackend` when file tools should read/write a path-keyed store instead of disk (`PowerMemPathStore` is one ready-made option)
 - **Get started in seconds** — `uv add deepagents` and you have a working agent
 - **Customize in minutes** — Add tools, swap models, tune prompts when you need to
 

examples/README.md

@@ -17,6 +17,7 @@
 | [deep_research](deep_research/) | Multi-step web research agent using Tavily for URL discovery, parallel sub-agents, and strategic reflection |
 | [content-builder-agent](content-builder-agent/) | Content writing agent that demonstrates memory (`AGENTS.md`), skills, and subagents for blog posts, LinkedIn posts, and tweets with generated images |
 | [text-to-sql-agent](text-to-sql-agent/) | Natural language to SQL agent with planning, skill-based workflows, and the Chinook demo database |
+| [memory-backend-agent](memory-backend-agent/) | Path-keyed **virtual files** (`MemoryBackend` + `PathMemoryStore`): same file tools, storage not on disk; in-memory demo plus optional **PowerMem** for persistence, multi-tenancy, and semantic search |
 | [ralph_mode](ralph_mode/) | Autonomous looping pattern that runs with fresh context each iteration, using the filesystem for persistence |
 | [downloading_agents](downloading_agents/) | Shows how agents are just folders—download a zip, unzip, and run |