strands-dakera

strands-dakera gives agents persistent memory that survives across sessions, backed by a self-hosted Dakera server. Unlike a fixed-TTL store, Dakera ranks recalled memories by importance × recency × semantic relevance, so the most contextually useful memories surface first.

It offers two integration points against the same server:

DakeraMemoryStore — a MemoryStore that plugs into the agent loop via a MemoryManager (Strands ≥ 1.45), with automatic memory injection and extraction.
dakera_memory — a tool the model calls explicitly for full CRUD (store / retrieve / get / update / delete).

Installation

pip install "strands-dakera>=0.2.0"   # DakeraMemoryStore requires >=0.2.0 (Strands >=1.45)

Run a Dakera server locally (once):

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

Memory store

Wire Dakera into the agent loop with a MemoryManager. The manager searches the store to recall context (injected into the prompt automatically) and, when writable, writes new memories — either directly or by extracting facts from the conversation every few turns.

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

store = DakeraMemoryStore(agent_id="alex", writable=True, extraction=True)
agent = Agent(memory_manager=MemoryManager(stores=[store]))

# Recall and writes happen automatically — no explicit tool call required.
agent("Remember that I prefer dark-mode dashboards and async standups.")
agent("How do I like to work?")  # recalls the stored preferences

DakeraMemoryStore implements search (decay-weighted recall) and add (a client-side write sink), so enabling extraction uses the manager’s client-side ModelExtractor. Common arguments: agent_id (required), name, writable, extraction, max_search_results, importance, memory_type, and base_url / api_key.

Tool

Prefer an explicitly model-invoked tool? Register dakera_memory instead (or alongside the store):

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,
)

Key Features

Decay-weighted recall: results ranked by a decay-aware importance signal, not just vector distance
Importance-typed storage: importance (0.0–1.0) and memory_type (episodic / semantic / procedural / working)
Two integration points: a MemoryStore for the agent loop and a full-CRUD tool, sharing one server
Self-hosted: runs against your own Dakera server — no cloud key required
Safe by default: mutative tool actions prompt for confirmation unless BYPASS_TOOL_CONSENT=true

Configuration

DAKERA_BASE_URL=http://localhost:3000   # Dakera server URL (default)
DAKERA_API_KEY=dk-your-key              # Optional API key