strands-dakera

Persistent, decay-weighted memory for Strands Agents

A community Strands Agents integration backed by a self-hosted Dakera memory server.

---

strands-dakera gives Strands agents a dakera_memory tool for durable memory that survives across sessions. Unlike a fixed-TTL store, Dakera ranks recalled memories by importance × recency × semantic relevance, so the most contextually useful memories surface first.

• Decay-weighted recall — retrieve returns memories ranked by a decay-aware importance signal, not just vector distance.
• Importance-typed storage — store accepts an importance (0.0–1.0) and a memory_type (episodic / semantic / procedural / working).
• Self-hosted, no cloud key — runs against a local dakera-ai/dakera-deploy docker-compose stack (server + MinIO, port 3000).

Install

pip install strands-dakera

Run a Dakera server (once):

git clone https://github.com/dakera-ai/dakera-deploy && cd dakera-deploy && docker compose up -d

Configure via environment variables:

Variable	Default	Description
DAKERA_BASE_URL	http://localhost:3000	Dakera server URL
DAKERA_API_KEY	(unset)	API key for the Dakera server (dk-...)

Usage

from strands import Agent
from strands_dakera import dakera_memory

agent = Agent(tools=[dakera_memory])

# Store a memory with an importance weight
agent.tool.dakera_memory(
    action="store",
    agent_id="alex",
    content="Alex prefers dark-mode dashboards and async standups.",
    importance=0.8,
    metadata={"category": "preferences"},
)

# Decay-weighted semantic recall
agent.tool.dakera_memory(
    action="retrieve",
    agent_id="alex",
    query="how does alex like to work?",
    top_k=5,
)

Actions

Action	Required fields	Description
store	agent_id, content	Store a new memory (importance, memory_type, metadata optional)
retrieve	agent_id, query	Decay-weighted semantic search (top_k optional, default 5)
get	agent_id, memory_id	Fetch a specific memory by ID
update	agent_id, memory_id	Update content / metadata of an existing memory
delete	agent_id, memory_id	Delete a memory by ID

Mutative actions (store, update, delete) prompt for confirmation unless BYPASS_TOOL_CONSENT=true.

See docs/dakera_memory_tool.md for the full action reference.

Memory store

DakeraMemoryStore is a Strands MemoryStore — a MemoryManager extension point (Strands ≥ 1.45). Where the dakera_memory tool is called explicitly by the model, a store plugs into the agent loop directly: the manager searches it to recall context (injected into the prompt automatically) and, when writable, writes new memories — either directly or via periodic extraction from the conversation. Both share one Dakera server and agent namespace.

from strands import Agent
from strands.memory import MemoryManager
from strands_dakera import DakeraMemoryStore

# Recall + write, distilling facts from the conversation every few turns.
store = DakeraMemoryStore(agent_id="alex", writable=True, extraction=True)
agent = Agent(memory_manager=MemoryManager(stores=[store]))

# The agent now recalls from and writes to Dakera without any explicit tool call.
agent("Remember that I prefer dark-mode dashboards.")
agent("How do I like my dashboards?")  # recalls the stored preference

DakeraMemoryStore implements search (decay-weighted recall) and add (a client-side write sink), so enabling extraction uses the manager's client-side ModelExtractor to distill facts before storing them.

Argument	Default	Description
agent_id	(required)	Dakera agent namespace that owns the memories
name	"dakera"	Store identifier, used to target it from memory tools
writable	True	Whether the manager may write to the store
extraction	None	Automatic extraction (bool or ExtractionConfig)
max_search_results	None	Default result cap per search (falls back to 5)
importance	None	Default importance (0.0–1.0) applied to writes
memory_type	"episodic"	Default Dakera memory type for writes
base_url / api_key	env	Override DAKERA_BASE_URL / DAKERA_API_KEY

Development

The package lives under python/ (monorepo-style layout matching the Strands extension-template).

cd python
pip install hatch
hatch run test        # pytest (no live server required — mocked client)
hatch run prepare     # format + lint + typecheck + test

Roadmap

• DakeraSessionManager — a Strands SessionManager that persists conversation state to Dakera as a first-class memory store (tracked separately).

License

Apache-2.0. Dakera is a trademark of Dakera. Strands Agents is a project of its respective authors.