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 →

What are you replacingor evaluating?

Pick the product you are looking at. Each page takes a different angle, because a managed memory API, a graph framework and a vector database are not the same decision.

Reviewed October 2026

A rack of four services (vector store, embedding service, knowledge graph, session store) next to a single Dakera blockA rack of four services (vector store, embedding service, knowledge graph, session store) next to a single Dakera block

A memory product

Purpose-built for agent memory: storing, recalling and managing context over time.

Mem0Your memory, your bill, your servers. Estimate your monthly cost.Read the comparison ZepZep Community Edition is retired. The self-hosted paths from here.Read the comparison LettaAn agent harness vs a memory layer you can put under any agent.Read the comparison CogneeWho builds your graph: an LLM pipeline or an on-device model?Read the comparison LangMemA library inside one LangGraph process vs a memory service every agent can share.Read the comparison SupermemoryA context platform with connectors vs a memory engine you own.Read the comparison HindsightWhat runs in production: Postgres and an LLM provider, or one binary.Read the comparison HonchoLLM-reasoned user models or a deterministic, on-device memory engine.Read the comparison

A vector database

General-purpose similarity search. Memory semantics such as decay, sessions and graphs are layers you add on top.

PineconeA vector database is a component; agent memory is a system. ChromaDBFrom notebook to production memory. QdrantRust and Rust: a great vector engine vs a complete memory engine. WeaviateModules you assemble vs an engine that ships integrated. MilvusBuilt for billions of vectors vs built for agent memory: right-size your stack. pgvectorYou already run Postgres. Here is what you would still write. RedisRAM-priced speed vs disk-backed memory with semantics.

What every comparison checks

The same five questions, asked of every product, so the pages can be read side by side.

Cost and licenceWhat it costs at your volume, and what you are allowed to run.
Where data livesWhich services see your memory text, and who holds the storage.
What runs in the memory pathWhether an LLM or external API is called on every store and recall.
Retrieval and rankingHow results are found, ranked and explained.
OperationsWhat you deploy, upgrade, back up and roll back.

Still deciding

Guides for teams evaluating Mem0 or comparing the field.

Mem0 alternativesA buyer’s guide: pick by what you need.Read the guide Mem0 self-hostingCount the moving parts.Read the guide

New in Dakera v0.12.0

Released 2026-10-01. Every feature is opt-in, and upgrades from v0.11.108 run in place.

What’s new in the docs
100+languages in one embedding space, plus speech and document-page recall
~2×faster reranked recall on the same CPU
7.6×less container memory after reranking (0.65 GB, was 4.98 GB)
1 commandto roll back: dakera downgrade
The Dakera crates stacked on a base that reads deploy anywhere: on premises, Kubernetes or any cloud.The Dakera crates stacked on a base that reads deploy anywhere: on premises, Kubernetes or any cloud.

Run Dakera on your own infrastructure

Self-hosted is free. One binary with its models baked in and no per-call fees. Start it with one Docker command and see it for yourself.

Run locally in 5 minutes Read the docs

Prefer managed hosting? Join the Dakera Cloud waitlist

Further reading: Best AI Agent Memory Frameworks 2026Benchmark methodologyHow Agent Memory Works