Wisteria
BYOK chatbot that saves thoughts and memory locally
About
Wisteria is a desktop chatbot that keeps conversations and context locally using a custom memory compression architecture called Anchoring. It preserves temporal context and subjective perspective while avoiding provider‑side reasoning, allowing users to inspect logs and reasoning in plain text. The bot supports popular LLMs and stores all data in the user’s own directory.
The case
Who it is for
Privacy‑conscious solo developers building personal AI assistants who want local memory control.
What is broken today
Large language models lack reliable long‑term memory, often repeating or hallucinating past events. Users cannot control where conversation data is stored, and provider‑side reasoning obscures how responses are generated. This limits privacy and the ability to build personal AI assistants that remember meaningful context over time.
What it does about it
Wisteria collects each turn of dialogue, then compresses it into memory chunks using a custom multi‑level abstraction. These chunks are stored locally as plain text files. A response schema injects the compressed memory back into the LLM prompt, enabling the model to recall prior context without relying on the provider’s internal memory. All logs, persona memories, and reasoning can be inspected by the user.
Why not the obvious alternative
Unlike generic LLM wrappers that rely on provider APIs for memory, Wisteria keeps all data on the user’s machine, preserving privacy. Its Anchoring compression preserves temporal ordering and subjective perspective, reducing hallucinations compared to standard abstract compression. The custom response schema lets users see the model’s internal reasoning, and the system can be paused during compression to maintain data integrity.
Details
Questions
- Is Wisteria free?
- Wisteria is sold as a one‑time purchase on itch.io; there is no free version.
- Does it store data locally?
- Yes, all conversation logs, memory chunks, and persona data are saved as plain text files in a folder on the user’s device.
- Which LLMs does it support?
- The client is designed to work with popular provider models that accept custom prompts; you can specify the model in the configuration.
- How long does memory compression take?
- Compression occurs in the background and may take a few seconds to several minutes depending on the model, session length, and compression level.
- Can I inspect the reasoning the model uses?
- Yes, the client logs the model’s raw responses and the reasoning text, which can be reviewed directly.
No comments yet. Be the first to give the maker feedback.