PLML
A user-controlled memory architecture proposal for AI assistants that addresses unreliable cross-session memory by storing conversation history locally, retrieving relevant fragments locally and sharing only necessary context with the model.
Timeline (last 90 days)
Bert Singels is a 54-year-old autistic Dutch man who, based on his own frustrating experience using AI assistants, proposed the PLML (Persistent Local Memory Layer) technical proposal.
The core principles of PLML are local storage, local retrieval, sharing only what is necessary, and decoupling the three functions of long-term storage, context retrieval, and language model inference.
The PLML proposal points out that the problem is not whether conversation data is stored, but whether the right information can be retrieved and supplied to the model at the right moment.
PLML uses SQLite or a similar local database to store conversation messages and their metadata, and uses full-text indexes and optional vector indexes to support retrieval.
PLML's local retrieval engine combines multiple methods such as full-text search, semantic search, metadata filtering, recency and relevance ranking, and source message retrieval.
The PLML design supports two integration modes simultaneously: Mode A (a browser extension interacting with existing chat interfaces) and Mode B (sending context directly via API or application integration), and does not bind the archive to either mode.
In scenarios where accuracy is critical, the context manager should prefer original source material; summaries can speed up retrieval but should not become the sole record of important technical details.
A key security stance of the PLML architecture is that archived messages should be treated as historical information rather than automatically trusted instructions; the local agent must not allow retrieved content to override current user intent or security controls.
The PLML proposal estimates that assuming a user exchanges about 24,000 words per day, at 6 bytes per word, a year of raw text is about 52MB; planning at 200MB per user per year gives 200TB for 1 million users and 180PB for 900 million users.
The PLML proposal advocates that local archives use open, documented data formats, so that users can switch between AI providers without having to rebuild their personal history or depend on proprietary memory formats.
1 more timeline events
All Facts (11)
Bert Singels is a 54-year-old autistic Dutch man who, based on his own frustrating experience using AI assistants, proposed the PLML (Persistent Local Memory Layer) technical proposal.
50%UnverifiedThe core principles of PLML are local storage, local retrieval, sharing only what is necessary, and decoupling the three functions of long-term storage, context retrieval, and language model inference.
50%UnverifiedThe PLML proposal points out that the problem is not whether conversation data is stored, but whether the right information can be retrieved and supplied to the model at the right moment.
50%UnverifiedPLML uses SQLite or a similar local database to store conversation messages and their metadata, and uses full-text indexes and optional vector indexes to support retrieval.
50%UnverifiedPLML's local retrieval engine combines multiple methods such as full-text search, semantic search, metadata filtering, recency and relevance ranking, and source message retrieval.
50%UnverifiedThe PLML design supports two integration modes simultaneously: Mode A (a browser extension interacting with existing chat interfaces) and Mode B (sending context directly via API or application integration), and does not bind the archive to either mode.
50%UnverifiedIn scenarios where accuracy is critical, the context manager should prefer original source material; summaries can speed up retrieval but should not become the sole record of important technical details.
50%UnverifiedA key security stance of the PLML architecture is that archived messages should be treated as historical information rather than automatically trusted instructions; the local agent must not allow retrieved content to override current user intent or security controls.
50%UnverifiedThe PLML proposal estimates that assuming a user exchanges about 24,000 words per day, at 6 bytes per word, a year of raw text is about 52MB; planning at 200MB per user per year gives 200TB for 1 million users and 180PB for 900 million users.
50%UnverifiedThe PLML proposal advocates that local archives use open, documented data formats, so that users can switch between AI providers without having to rebuild their personal history or depend on proprietary memory formats.
50%UnverifiedPLML's development roadmap is divided into six incremental phases: local archive, automatic capture, context retrieval, browser automation, multi-provider support, and reliability and security.
50%