NewRelease 0.12.3 is out: Engine v0.12.3, Dashboard 0.5.2, Helm 0.12.3, the four SDKs, CLI 0.8.2 and MCP 0.12.3Release 0.12.3 is outSee what's new →
PUBLIC ALPHA

Dakera documentation

Dakera is a self-hosted memory engine for AI agents: one Rust server that stores agent memories and documents, embeds them on the machine it runs on, and recalls them with vector, full-text and hybrid search, ranking and a knowledge graph. It is served over REST (port 3000) and gRPC (port 50051), with SDKs for Python, TypeScript, Go and Rust, the dk CLI and an MCP server.

These pages document v0.12.3. Start with the Quick Start if you are new to Dakera, or with the upgrade guide if you run v0.11.108, v0.12.0, v0.12.1 or v0.12.2; the table below points to the page for each task.

5 minutesQuick StartRun the server with Docker, store three memories and recall them by meaning. No installPlaygroundCall the live API from your browser before you run anything. Claude · Cursor · WindsurfMCP serverGive the coding agents you already use a memory that persists. ReferenceREST APIEvery route, field and response, with examples in each SDK.
SDKsPythonTypeScriptGoRustCLI
Public alpha — the self-hosted server, SDKs, CLI and MCP server are available now. Dakera Cloud (managed hosting) is coming next — join the waitlist →
Dakera v0.12.3 — You are reading the latest docs. v0.12.3 is a patch release: agents import PDF, DOCX, HTML, Markdown and plain text as searchable, citable memory; recall results carry a rank_score that follows their order and neighborhood_mode places session context; event_time dates imported histories; ingest and reranked recall are faster; the images are signed and ship without curl (dakera healthcheck) (changelog →). v0.12.2 added key prefix patterns, in-place key edits and rotation grace, POST /v1/agents and idle session auto-end. v0.12.1 made DAKERA_CLUSTER_SECRET required on every cluster node. New in v0.12: opt-in multimodal memory (attachments, speech to text, image indexing, records), late interaction, multilingual search and RaBitQ; GET /v1/capabilities; gRPC authentication; enforced namespace quotas; and dakera downgrade to go back. What's new in v0.12 → · Upgrading to v0.12 → · Rolling back to v0.11 → · Looking for the previous release? v0.11 docs →

Find the page for a task

I want to…Go to
Upgrade from v0.11.108, v0.12.0, v0.12.1 or v0.12.2 (or roll back)Upgrading → · Rolling back →
See everything v0.12.0 introducesWhat's new → · Release overview →
Search in several languagesMultilingual search →
Import PDFs, Word files, web pages and Markdown as citable memoryDocument import →
Store audio and page images as memoriesMultimodal memory → · Records & multi-vector →
Use late interaction, RaBitQ or rerank controlsSearch & ranking →
Run models offline, in Docker or KubernetesModels & Docker →
Operate a deployment (health, backups, rollback)Operations →
Operate a server from a web UIDashboard →
See the measured performance numbersPerformance →
Get Dakera running in 5 minutesQuick Start →
Add memory to Claude Desktop, Code, or CursorMCP Server →
Integrate from PythonPython SDK →
Integrate from TypeScript / NodeTypeScript SDK →
Integrate from Go or RustGo SDK → · Rust SDK →
Use from the command lineCLI Reference →
See all configuration optionsConfiguration →
Deploy to productionDeployment →
Understand the retrieval pipelineArchitecture →
Security hardening for productionSecurity →
Multi-node clusteringHigh Availability →
Use with LangChain, CrewAI, LlamaIndex, or AutoGenIntegrations →
Workflow of storing a memory. Your agent sends content with POST /v1/memory/store. Dakera validates the request (auth, key scope, quotas) and refuses a write over a quota with 413; it embeds the content on-device with ONNX, writes it to the write-ahead log and storage, indexes it in HNSW and BM25, and links it in the knowledge graph to similar memories and memories that share an entity. The agent gets back the stored memory with its id, importance, tags and embedding_time_ms.Workflow of storing a memory. Your agent sends content with POST /v1/memory/store. Dakera validates the request (auth, key scope, quotas) and refuses a write over a quota with 413; it embeds the content on-device with ONNX, writes it to the write-ahead log and storage, indexes it in HNSW and BM25, and links it in the knowledge graph to similar memories and memories that share an entity. The agent gets back the stored memory with its id, importance, tags and embedding_time_ms.
Where a stored memory goes, from your agent to the indexes and the graph.Full size ↗Full size ↗

What Dakera replaces

A rack of four services (vector store, embedding service, knowledge graph, session store) next to a single Dakera block
Typical stack · 3–5 services, 1–2 GB  ·  Dakera · 1 binary, ~101 MB
Instead of runningDakera provides
Qdrant · Pinecone · WeaviateHNSW vector index with SIMD-accelerated distances, plus an opt-in RaBitQ search mode
Elasticsearch · OpenSearchBM25 full-text search engine with per-namespace indexes
OpenAI / Cohere embeddings APIOn-device ONNX inference (MiniLM, BGE, E5; opt-in bge-m3 multilingual and colbert-small late interaction) — zero API calls
Redis / Postgres memory layerDecay-weighted agent memory with sessions, importance scoring, and 6 decay strategies
Neo4j knowledge graphMemory graph with 4 edge types (related, shared entity, precedes, explicit link) and a cross-agent network
Mem0 / Zep memory servicesImport and export in Mem0 and Zep formats (and JSONL, CSV)
Separate NER serviceGLiNER zero-shot named entity extraction with multi-provider support
88.2% Recall@20 on LoCoMo — Dakera v0.11.107 scored 88.2% Recall@20 on the LoCoMo benchmark (10 conversations, 1,536 evaluated questions); this figure was measured on v0.11.107, not on v0.12.0 (v0.12.0 release measurements are in What's new). Read the methodology →

Key capabilities

CapabilityDescription
Hybrid RetrievalVector ANN and BM25 full-text combined (min-max or Reciprocal Rank Fusion), then cross-encoder reranking. HNSW search p50 0.87 ms on BEIR Quora 50k; see Performance.
On-device InferenceEmbeddings (MiniLM, BGE, E5), reranking, and entity extraction via ONNX Runtime — zero external API calls.
Memory DecaySix decay strategies. Per-type TTLs. Spaced repetition. Configurable half-life.
Knowledge GraphsMemory graph with 4 edge types, BFS traversal, shortest paths and a cross-agent network.
HA ClusteringGossip membership and leader election; every node holds a full copy and accepts writes, and versioned replication (hybrid logical clock, tombstones) converges all nodes. Details →
Multilingual SearchOpt-in bge-m3 embeddings, per-language full-text stemming and stop words, CJK bigram indexing, dates understood in seven query languages, and a per-request lang. Details →
Multimodal MemoryOpt-in attachments, speech to text, image-page indexing with visual recall, and multi-vector records. Every media job reserves its memory first. Details →
Late Interaction & RaBitQOpt-in per-token MaxSim reranking (colbert-small) and a RaBitQ search mode for latency. Details →
One-command Rollbackdakera downgrade converts a stopped deployment's data back to what v0.11.108 reads. Details →
Enterprise SecurityAES-256-GCM encryption at rest with key rotation, scoped API keys, rate limiting, input validation against injection patterns, audit logging.
AutoPilotBackground deduplication and consolidation of low-importance memories.
MCP Native14 core tools (102 in total across profiles) for Claude Desktop, Claude Code, Cursor and Windsurf, with no code to write. Details →
Dual ProtocolREST API on port 3000 and gRPC API on port 50051. Both share the same engine, storage, and auth layers.
Query ClassificationEach query is classified (keyword, semantic, hybrid, temporal, multi-hop) and routed to a matching retrieval strategy. Rule-based by default; an embedding-based classifier is opt-in (DAKERA_ML_CLASSIFIER).
Session ManagementGroup memories by agent session, recall within one session, and end a session with a summary (written by you or generated by the server).
Entity ExtractionGLiNER zero-shot NER with a rule-based pre-pass (dates, URLs, UUIDs, emails, IPs). Other extraction providers (OpenAI, Anthropic, OpenRouter, Ollama) are configurable per namespace.
Backup & RecoveryScheduled backups, compressed (zstd) or encrypted. Write-ahead log and snapshots. Import/export (Mem0, Zep, JSONL, CSV).
Cross-encoder Rerankingbge-reranker-v2-m3 ONNX model (shipped in the image) for relevance scoring. Applied after initial retrieval for higher-precision results.

Packages

Every component of the current release train, with its install command. Versions come from the same data as the changelog.

Current versions — release train 0.12.3 (Oct 22, 2026)
ComponentVersionReleasedInstallLinks
Enginedakera 0.12.3 docker pull ghcr.io/dakera-ai/dakera:0.12.3 Release notesGHCR
Dashboarddakera-dashboard 0.5.2 docker pull ghcr.io/dakera-ai/dakera-dashboard:0.5.2 Dashboard docsGHCR
Helm chartdakera-helm 0.12.3 helm install dakera oci://ghcr.io/dakera-ai/dakera-helm/dakera --version 0.12.3 GitHub releaseArtifact HubCHANGELOG
Python SDKdakera 0.14.1 pip install dakera==0.14.1 GitHub releasePyPICHANGELOG
TypeScript SDK@dakera-ai/dakera 0.12.3 npm install @dakera-ai/dakera@0.12.3 GitHub releasenpmCHANGELOG
Go SDKgithub.com/dakera-ai/dakera-go v0.12.3 go get github.com/dakera-ai/dakera-go@v0.12.3 GitHub releasepkg.go.devCHANGELOG
Rust SDKdakera-client 0.12.3 cargo add dakera-client@0.12.3 GitHub releasecrates.ioCHANGELOG
CLIdakera-cli 0.8.2 brew install dakera-ai/tap/dk GitHub releasecrates.ioCHANGELOG
MCP server@dakera-ai/dakera-mcp 0.12.3 npm install -g @dakera-ai/dakera-mcp@0.12.3 GitHub releasenpmCHANGELOG
Framework integrations (released on their own schedule)
langchain-dakeraPython 3.10+ 0.3.0 — pip install langchain-dakera PyPI
llamaindex-dakeraPython 3.10+ 0.3.0 — pip install llamaindex-dakera PyPI
crewai-dakeraPython 3.10+ 0.3.0 — pip install crewai-dakera PyPI
autogen-dakeraPython 3.10+ 0.3.0 — pip install autogen-dakera PyPI
strands-dakeraPython 3.10+ 0.3.0 — pip install strands-dakera PyPI
@dakera-ai/langchainNode.js 20+ 0.3.0 — npm install @dakera-ai/langchain npm
@dakera-ai/ai-sdkNode.js 20+ 0.2.0 — npm install @dakera-ai/ai-sdk npm
Dakera sends product telemetry by default. Since v0.12 it includes the machine's hostname, and PostHog records the connecting IP (GeoIP). Opt out: DAKERA_TELEMETRY=0 or DO_NOT_TRACK=1. Details →

Release artifacts

Where each released component is published.

Stay sharp on agent memory
Benchmark results, SDK releases, and production patterns. Under 500 words per issue.
✓ You're in. First issue lands soon — watch for Dakera in your inbox.