Plast-Mem

Why it's ranked here

The design is unusually explicit about memory lifecycle, from ingestion through segmentation, consolidation, retrieval, and review. The current implementation also exposes meaningful gaps: core functionality remains incomplete, surprise-based retention is not active, and some accepted API options have no effect. It is promising infrastructure, not a finished component.

What's good

Episodic and semantic records have distinct retention rules instead of sharing one vague memory model. PostgreSQL-backed jobs separate ingestion from expensive model work. Retrieval mixes lexical and semantic signals, while episodic results also account for modeled retrievability. Semantic facts preserve provenance and use invalidation rather than hard deletion.

Tradeoffs

Deployment requires PostgreSQL with ParadeDB plus compatible chat and embedding services. The server port is fixed. Semantic facts have no direct write interface. Graph retrieval is only future design work. Current episode creation sets surprise to zero, so the documented surprise-based stability boost is absent. The retrieval detail option is accepted but ignored by formatting.

How to use it well

Use it when building a self-hosted conversational agent that needs asynchronous memory extraction and hybrid recall. Treat the HTTP service as a memory subsystem beside your model orchestration, and validate its unfinished behavior before production use. It does not provide a graph memory system, a hosted service, or general agent orchestration.

Observed

License

MIT

Primary language

Rust

Packaging

Single binary or Docker image

Service interface

HTTP API with an OpenAPI surface

Client interface

Generated TypeScript client package

Runtime dependencies

PostgreSQL with ParadeDB and OpenAI-compatible chat and embedding services

Repository structure

Rust workspace with ten members plus a private pnpm workspace