Dakera vs Hindsight: what you run in production
Hindsight is an MIT-licensed memory system from Vectorize with retain, recall and reflect operations, banks and an LLM-built memory model. Both can be self-hosted. This page is about the day-two question: what you operate, patch, back up and roll back once an agent depends on it.
Reviewed October 2026
What you run
Hindsight needs PostgreSQL 14+ with a vector extension and an LLM provider for fact extraction, entity resolution and answer generation. Its Docker image embeds PostgreSQL by default and can use an external one; the Helm chart expects PostgreSQL as its backing store. Dakera is one container whose models are in the image.
docker run -it --name hindsight \
--restart unless-stopped \
-p 8888:8888 -p 9999:9999 \
-e HINDSIGHT_API_LLM_API_KEY=$KEY \
-v hindsight-data:/home/hindsight/.pg0 \
ghcr.io/vectorize-io/hindsight:latestdocker run -d --name dakera \
--restart unless-stopped \
-p 3000:3000 \
-e DAKERA_ROOT_API_KEY=my-dev-key \
-e DAKERA_STORAGE=filesystem \
-v dakera-data:/data \
-v dakera-models:/app/models \
ghcr.io/dakera-ai/dakera:0.12.0
curl localhost:3000/health/readyBoth start a working server in one command. The difference shows up afterwards, in the checklist below.
The operations checklist
What an on-call engineer asks about, and where each answer lives.
| Capability | Hindsight | Dakera v0.12.0 |
|---|---|---|
Liveness and readiness probesDakera: /health/live and /health/ready, so a model download is not mistaken for a hang | Yes | Yes |
Metrics and dashboardsHindsight: /metrics, Grafana dashboards and OpenTelemetry traces. Dakera: Prometheus alert rules and a Grafana dashboard in dakera-deploy | Yes | Yes |
| BackupsHindsight: through your PostgreSQL tooling. Dakera: encrypted, compressed bundles through the backup API, restorable with their attachments | PostgreSQL tooling | Built in |
| Start with no networkHindsight can pair with a local LLM provider; Dakera ships its default models in the image, ready about 4.4 s after start with no download | Partly | Yes |
| Encryption at restAES-256-GCM once a key is configured, bound to its record and namespace, with a replicated keyring and per-namespace rotation | PostgreSQL or disk encryption | Built in |
dakera --check-config validates configuration before anything starts. Upgrades from v0.11.108 run in place, and dakera downgrade converts data back if you need to roll back. dakera models list|pull|prune pre-seeds models for air-gapped hosts.Where the work happens
Hindsight's retain step uses your LLM provider to extract facts, so memory writes cost tokens; its Cloud bills retain at $10 per million tokens, recall at $0.75 per million and reflect at $0.05 per call. In Dakera, store and recall are deterministic and on-device, so the cost is the server you already run. Teams that want an LLM to reason over memory, as Hindsight's reflect does, can still call their model of choice from the agent, with Dakera supplying the recalled memories.
For fleets, Dakera adds multi-node high availability with replicated writes and a durable outbox that merges instead of overwriting, plus authenticated, rate-limited, audited gRPC for vector operations. The Dashboard gives an overview and health page, a per-agent memory browser and a Recall lab. For the full runbook see Operations and Rolling back.
Feature notes
| Hindsight | Dakera | |
|---|---|---|
| Memory model | LLM-extracted facts, entities and relationships in banks, with retain, recall and reflect | Memories with on-device entity extraction, a graph with 5 edge types, 6 decay curves, consolidation and deduplication |
| Retrieval | Four parallel strategies merged with RRF and a cross-encoder | HNSW plus BM25 with RRF, then a cross-encoder rerank with per-request rerank_candidates |
| Languages and media | Not covered in our review | Multilingual embeddings and stemming, speech to text, visual recall over document pages |
| Clients | Python, Node, Go and a CLI; MCP endpoint per bank | Python, TypeScript, Go and Rust SDKs; MCP with 14 core tools (90+ via profiles) |
| Licence | MIT, server included; managed Hindsight Cloud available | SDKs, CLI and MCP server are MIT; the server binary is proprietary with no usage fees |
| LoCoMo | Publishes accuracy results; see its benchmark site | 88.2% Recall@20 asks whether retrieval surfaces the answer, while accuracy asks whether an LLM answers correctly. Different measurements. Method |
When to choose each
Choose Hindsight if
- You want the whole server under an MIT licence, which Dakera does not offer.
- You already run PostgreSQL and an LLM provider, so they are not extra parts for you.
- You value LLM-built observations and a reflect operation that reasons over stored memory.
- You want a managed cloud with pure usage-based pricing and no seat fees.
Choose Dakera if
- You want one artifact to deploy, monitor, back up and roll back.
- The host is air-gapped or memory text must not reach a model provider.
- You want fixed infrastructure cost instead of token-metered writes.
- You need multilingual or multimodal memory on your own hardware.
Prefer managed? Dakera Cloud is coming. Join the waitlist.
Frequently asked questions
Does Hindsight need an LLM?
Per its installation guide, yes: an LLM provider key is needed for fact extraction, entity resolution and answer generation. Providers include hosted APIs and local options such as Ollama.
Does Hindsight need PostgreSQL?
Yes, PostgreSQL 14+ with a vector extension. The Docker image embeds one by default; it also supports an external PostgreSQL and Oracle AI Database 23ai.
What does Dakera need to run?
A single container and one secret protecting your own server, DAKERA_ROOT_API_KEY. No database, LLM or embedding provider is required.
Can I go back after upgrading Dakera?
Yes. dakera downgrade converts data back to what v0.11.108 reads, and upgrades from v0.11.108 run in place.
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 →