Dakera AI

One self-hosted engine for storing and recalling what your agents learn: hybrid vector, full-text and graph recall,
embeddings and reranking built in, on CPU, with no LLM call at recall time.

Website  ·  Docs  ·  Quickstart  ·  Benchmark  ·  What's new in v0.12.0

Dakera is a self-hosted memory engine for AI agents. Agents store what they learn and recall it later by meaning, keyword, time and relationship, across sessions and across agents.

It ships as one container image with its models inside: embeddings, a cross-encoder reranker and entity extraction run locally on CPU. Recall never calls an LLM, so cost and latency stay predictable, and your data never leaves your infrastructure.

Under the hood: HNSW and IVF vector indexes, BM25 full text, a knowledge graph, sessions, namespaces and decay-weighted importance, served over REST, gRPC and MCP.

_{CPU-only figures, v0.12.0 compared with v0.11.108 on the same machine: vector search and index memory on BEIR Quora, 50k 1024-d vectors; reranking with bge-reranker-v2-m3 at top_k 16 in a 4 vCPU / 8 GiB container; multimodal stack = working memory with every model loaded. Details in the release notes.}

1. Run the server. A fresh install with authentication on needs a root API key. The image ships its default models, so it starts without a download.

export DAKERA_API_KEY="dk-$(openssl rand -hex 16)"

docker run -d --name dakera -p 3000:3000 \
  -v dakera-data:/data \
  -v dakera-models:/app/models \
  -e DAKERA_ROOT_API_KEY="$DAKERA_API_KEY" \
  -e DAKERA_STORAGE=filesystem \
  ghcr.io/dakera-ai/dakera:0.12.0

curl -s http://localhost:3000/health/ready   # 200 once the models are loaded

2. Store a memory, then recall it.

curl -s http://localhost:3000/v1/memory/store \
  -H "Authorization: Bearer $DAKERA_API_KEY" -H "Content-Type: application/json" \
  -d '{"agent_id": "my-agent", "importance": 0.8,
       "content": "The user prefers concise answers with code examples"}'

curl -s http://localhost:3000/v1/memory/recall \
  -H "Authorization: Bearer $DAKERA_API_KEY" -H "Content-Type: application/json" \
  -d '{"agent_id": "my-agent", "query": "How does the user like answers?", "top_k": 5}'

import os
from dakera import DakeraClient

client = DakeraClient(base_url="http://localhost:3000", api_key=os.environ["DAKERA_API_KEY"])

client.store_memory(
    agent_id="my-agent",
    content="The user prefers concise answers with code examples",
    importance=0.8,
    tags=["preference"],
)

response = client.recall(agent_id="my-agent", query="How does the user like answers?", top_k=5)
for memory in response.memories:
    print(f"{memory.score:.2f}  {memory.content}")

import { DakeraClient } from '@dakera-ai/dakera';

const client = new DakeraClient({ baseUrl: 'http://localhost:3000', apiKey: process.env.DAKERA_API_KEY });

await client.storeMemory('my-agent', {
  content: 'The user prefers concise answers with code examples',
  importance: 0.8,
  tags: ['preference'],
});

const { memories } = await client.recall('my-agent', 'How does the user like answers?', { top_k: 5 });
memories.forEach((m) => console.log(m.score.toFixed(2), m.content));

_{The snippets use the released SDKs (dakera 0.12.12 on PyPI, @dakera-ai/dakera 0.11.106 on npm), which cover these calls on a v0.12.0 server. Client releases with the v0.12 additions are coming. Pin an image version in production; :latest tracks the newest release.}

_{Recall@20 on the LoCoMo evaluation set — 10 conversations, 1,536 evaluated questions (adversarial category excluded), no LLM in the retrieval path. Dakera v0.11.107.}

Methodology and results: dakera.ai/benchmark

The Dakera server is distributed as a container image, ghcr.io/dakera-ai/dakera. Everything else below is open source.

# Docker Compose: Dakera with MinIO object storage
git clone https://github.com/Dakera-AI/dakera-deploy && cd dakera-deploy/docker
cp .env.example .env          # set DAKERA_ROOT_API_KEY and the MinIO credentials
docker compose up -d

# Kubernetes: the Helm chart
helm install dakera oci://ghcr.io/dakera-ai/dakera-helm/dakera \
  --namespace dakera --create-namespace \
  --set dakera.rootApiKey="$DAKERA_API_KEY" \
  --set minio.rootPassword="<password>"

Configuration, clustering, storage backends and air-gapped installs: deployment docs.

• Questions and bugs: open an issue on the repository you use, for example dakera-py or dakera-deploy. Include the server version from /health. See SUPPORT.md.
• Security: report vulnerabilities privately, never in a public issue. See SECURITY.md.
• Contributing: the SDKs, integrations, CLI, MCP server and deployment repositories accept pull requests. See CONTRIBUTING.md and the code of conduct.
• News: LinkedIn and the blog.

The SDKs, framework integrations, CLI and MCP server are MIT licensed (strands-dakera is Apache-2.0); check each repository. The Dakera server is proprietary and is distributed as a container image with no usage fees.

_{The server sends product telemetry, never memory content, unless you turn it off with DAKERA_TELEMETRY=0 or DO_NOT_TRACK=1. What it sends is listed in the telemetry docs.}

dakera.ai  ·  Docs  ·  Benchmark  ·  LinkedIn
_{ذاكرة · dhākira · Arabic for memory}