Dakera vs Redis: RAM-priced speed vs disk-backed memory with semantics
Redis is an in-memory data store with a built-in query engine, vector sets and, through Redis Cloud, a managed agent memory service. Dakera keeps memories on disk, bounds what it holds in memory, and adds memory semantics on top. The trade is where the bytes live, and what each memory costs to keep.
Reviewed October 2026
Where the bytes live
Redis describes itself as an in-memory data store, with replication and on-disk persistence. That is the source of its speed, and it means the working set sits in RAM. Redis Flex, offered on Redis Cloud, splits data across RAM and SSD at an adjustable ratio for larger datasets.
Footprint, measured and documented
Redis can store vectors compactly: FLOAT16 and, since 8.0, INT8 types, plus SVS-VAMANA compression. Dakera's v0.12.0 work shrank its own index and its model working set.
Measured on BEIR Quora (50,000 vectors, recall@10 0.998 in both versions) and in the v0.12.0 reranking test (4 vCPU / 8 GiB container, bge-reranker-v2-m3). Dakera keeps cold data on disk or object storage through tiered storage, so RAM holds the working set, not the whole corpus.
Agent memory on each side
Redis announced Redis Iris in May 2026, a context layer built from five tools. Two matter here: Redis Agent Memory, a managed service on Redis Cloud, and LangCache, a semantic cache that Redis says can save up to 90 percent of token costs.
| Redis | Dakera | |
|---|---|---|
| Session and long-term memory | Agent Memory keeps session-scoped working memory and long-term memory, with REST, MCP and Python interfaces | Sessions, agents and namespaces in the server; 14 core MCP tools; Python, TypeScript, Go and Rust SDKs |
| How memories are extracted | Configurable LLM-based extraction policies, then embedded and stored in Redis | Embeddings, reranking and entity extraction run on-device: no LLM and no external API in the memory path |
| Lifecycle | Summarizes and trims session history, promotes facts to long-term memory, deduplicates by content hash | 6 decay curves, spaced-repetition TTL, deduplication, consolidation, and a knowledge graph with a dashboard Recall lab |
| Licence | Redis 8 adds AGPLv3 as a licence option; Redis Cloud is a managed service | SDKs, CLI and MCP server under MIT; the server is proprietary and self-hosted with no usage fees |
What v0.12.0 changes for a Redis buyer
- Cost shape: memory is paid for in RAM or in disk. A smaller index and a 0.65 GB working set let Dakera run on a modest server instead of a RAM-sized one.
- Resilience: replication is versioned and merges instead of overwriting, with a durable outbox and tombstones, and encryption at rest uses a replicated keyring with per-namespace rotation.
- Languages and media: bge-m3 embeddings across 100+ languages, attachments, on-device speech to text and visual recall over document pages.
- Operations:
/health/ready,dakera --check-config, Prometheus alert rules, a Grafana dashboard anddakera downgrade.
When to choose Redis, and when Dakera
Choose Redis if
- Redis already runs in your stack and the agent should use the same system, with its clients and tooling.
- You need in-memory latency for state that is read on every turn, such as voice or fraud scoring.
- You want active-active multi-region and high availability tiers from Redis Cloud.
- Semantic caching of LLM responses (LangCache) is part of the plan.
- You want Redis Cloud to run extraction, memory and retrieval as managed services.
Choose Dakera if
- Memories accumulate for months, and you do not want their size to set your RAM bill.
- You want extraction, reranking and embeddings to run on your hardware without an LLM call per memory.
- You need decay, consolidation and a knowledge graph as built-in behaviour.
- You want multilingual or multimodal memory in one self-hosted binary.
Redis Cloud has a free 30 MB database and paid Essentials and Pro plans. Prefer managed? Dakera Cloud is coming. Join the waitlist.
Frequently asked questions
Is Redis a vector database?
Yes. Redis includes a Query Engine with FLAT, HNSW and SVS-VAMANA vector indexes, metadata filters and hybrid text and vector queries, and vector sets as a data type introduced in Redis 8.
Does Redis have an agent memory product?
Yes. Redis Agent Memory, part of Redis Iris, is a managed memory layer on Redis Cloud, announced in preview in May 2026. Dakera is the self-hosted alternative with no LLM in the memory path.
Why does footprint matter for memory?
Agent memory grows with every conversation. In-memory stores price that growth in RAM; Dakera stores memories on disk and holds a bounded set of indexes in memory.
How should I read the LoCoMo figure?
Dakera reports Recall@20: whether the answer is among the top 20 retrieved memories, with no LLM in the retrieval path. It measures memory retrieval, so it is not set against Redis's latency figures.
Run Dakera on your own infrastructure.
One binary with its models baked in and no per-call fees. Start it with one Docker command and see it for yourself.
Prefer managed hosting? Join the Dakera Cloud waitlist →