Add Dakera — self-hosted AI agent memory server (open core)

[ferhimedamine]

Project: Dakera

Self-host / deploy: https://github.com/dakera-ai/dakera-deploy
Website: https://dakera.ai

What it is: Dakera is an open-core, self-hosted memory server for AI agents — a single binary that gives agents persistent, decay-weighted long-term memory across sessions. It exposes a REST API and an MCP server, with vector + hybrid (semantic + full-text) search, knowledge-graph associations, and session tracking.

Why it fits Awesome-LLMOps: It ships as production infrastructure — Docker Compose, Kubernetes, and Helm charts, with HA cluster setup and monitoring (see dakera-deploy). It sits alongside the existing Vector search / serving entries as the retrieval/memory layer for agentic LLM apps, and is self-hostable so teams keep their agent memory on their own infra.

Suggested section: Vector search (or a dedicated Memory entry if you prefer).

Happy to match whatever format/section fits your contribution guidelines — thanks for maintaining this list!

[kehansama]

Nice addition — self-hosted agent memory is exactly the direction the ecosystem needs to go. Teams are increasingly wary of sending agent memory (which often contains project secrets, architecture decisions, and user preferences) to third-party services.

A few observations from the agent memory space:

1. Memory as infrastructure, not just a feature
Agent memory is evolving from an app-level concern to an infrastructure primitive — similar to how caches and message queues became standard deployment components. Having it as a standalone, self-hostable service (like Dakera) is the right architectural choice.

2. The cross-agent challenge
The biggest gap right now is that most memory solutions are tightly coupled to a single agent framework. Teams running Claude Code + Codex + custom agents end up with fragmented memory silos. A protocol-agnostic memory layer that any agent can plug into would be a major unlock. This is the problem we are solving with AgentRelay — a universal memory relay that works across agent platforms with structured, project-scoped persistence.

3. Memory poisoning awareness
Recent research (e.g., the arXiv paper on memory poisoning attacks in LLM agents) highlights that persistent memory introduces new attack surfaces. Self-hosted solutions like Dakera can implement organizational security controls (access policies, audit logs, memory sanitization) that cloud services cannot easily offer.

4. Suggested related entries
If you are curating this list, you might also consider adding entries for:

• Memory protocol standards (emerging work on agent-to-memory APIs)
• Memory evaluation benchmarks (how to measure retrieval quality, freshness, relevance)
• Cross-agent memory synchronization patterns

Great work maintaining this resource — the LLMOps space needs more organized knowledge like this.