Persistent, local, encrypted cross-session memory for Google ADK agents — backed by Mimir.
| Backend | Dependencies | Encryption | Hybrid Search | Local |
|---|---|---|---|---|
| InMemoryMemoryService | None | ❌ | ❌ | ✅ |
| VertexAiMemoryBankService | GCP + Gemini | ❌ | Gemini-driven | ❌ |
| VertexAiRagMemoryService | GCP + RAG | ❌ | GCP vector | ❌ |
| MimirMemoryService | Single binary | ✅ AES-256 | ✅ BM25+FTS5+Dense | ✅ |
- Zero cloud dependencies — a single Rust binary, SQLite database, fully local
- AES-256-GCM encryption at rest — your memory data stays private
- Hybrid search — BM25 keyword + FTS5 + dense vector search
- 30+ MCP tools — remember, recall, synthesize, benchmark, federate, and more
- Ebbinghaus confidence decay — memories fade naturally, important ones persist
pip install adk-mimir-memoryThis package requires the mimir binary. Download it from:
https://github.com/Perseus-Computing-LLC/mimir/releases
Or build from source:
cargo install mimirfrom google.adk.agents import Agent
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from adk_mimir_memory import MimirMemoryService
agent = Agent(
name="my_agent",
model="gemini-2.5-flash",
instruction="You are a helpful assistant with persistent memory.",
)
# The memory service is configured on the Runner, not on the Agent.
runner = Runner(
agent=agent,
app_name="my_app",
session_service=InMemorySessionService(),
memory_service=MimirMemoryService(db_path="~/.adk/mimir.db"),
)That's it. Sessions, events, and explicit memories are now persisted across restarts.
# Custom database location
MimirMemoryService(db_path="/data/agent_memory.db")
# Custom mimir binary path (if not on $PATH)
MimirMemoryService(mimir_binary="/usr/local/bin/mimir")
# Both
MimirMemoryService(
db_path="/data/agent_memory.db",
mimir_binary="/usr/local/bin/mimir",
)This package also includes a drop-in agent with live workspace awareness via Perseus:
from google.adk.runners import Runner
from google.adk.sessions import InMemorySessionService
from adk_mimir_memory.perseus_context import perseus_context_agent
# The agent resolves @file, @search, @memory directives at inference time.
# Bind it on the Runner; run_async takes no `agent` argument.
runner = Runner(
agent=perseus_context_agent,
app_name="my_app",
session_service=InMemorySessionService(),
)
runner.run_async(
user_id="user",
session_id="session",
new_message=types.Content(role="user", parts=[types.Part.from_text(
text="What does the README say about deployment?"
)]),
)pip install adk-mimir-memory[perseus] # installs perseus-ctxSet directives via session state:
session = await runner.session_service.create_session(
app_name="my_app",
user_id="user",
state={
"_perseus_directives": "@file AGENTS.md @file README.md @memory deployment",
"_perseus_workspace": "/path/to/project",
},
)┌─────────────┐ JSON-RPC (MCP stdio) ┌──────────┐
│ ADK Agent │ ──────────────────────────▶ │ Mimir │
│ (Python) │ ◀────────────────────────── │ (Rust) │
└─────────────┘ └────┬─────┘
│
SQLite + FTS5
(AES-256-GCM)
The MimirMemoryService spawns a mimir subprocess and communicates via JSON-RPC over stdin/stdout (MCP stdio transport). Each add_session_to_memory, add_memory, and search_memory call translates to a Mimir MCP tool invocation.