AI Disclaimer

MemDev uses artificial intelligence and machine learning to compress terminal sessions into natural-language memories. This page explains how those systems work, their limitations, and your responsibilities when using AI-generated output.

How MemDev Uses AI

Two AI models run locally when you use MemDev:

• Working memory — Mamba SSM (state-spaces/mamba-130m): A frozen state-space model that ingests session text incrementally and maintains a fixed-size state vector capturing conversational flow, topic drift, and salience signals. This model is not trained online; it evolves deterministically during inference.
• Episodic compressor — CodeT5-Large (Salesforce/codet5-large, 770M params): Fine-tuned on 75,000 cleaned commit-message pairs from CommitBench. It converts raw terminal events and the SSM state into 1–2 sentence intent summaries stored as searchable memories.

No Guarantee of Accuracy

AI-generated session summaries are probabilistic outputs based on patterns learned from training data. They may be incomplete, imprecise, or occasionally misleading. MemDev does not guarantee that any compressed memory accurately represents what occurred in a terminal session.

You should treat MemDev memories as helpful context — not as authoritative logs or audit records.

Not a Verbatim Record

MemDev compresses sessions to reduce token usage and improve relevance. Summaries are interpretive, not verbatim. If you need an exact record of commands run, use a shell history tool or your terminal's scrollback buffer. MemDev is designed for semantic recall, not forensic accuracy.

Human Review Required

AI-generated context injected into your AI assistant conversations should be reviewed before you act on it. MemDev augments your workflow — it does not replace your judgment. Verify critical information from primary sources before making architectural, security, or production decisions based on MemDev output.

Not Professional Advice

Nothing generated by MemDev constitutes professional advice of any kind (legal, security, financial, or otherwise). Always consult qualified professionals for decisions in those domains.

Your Data, Your Models

MemDev does not use your terminal session data to train shared AI models. All model inference runs locally on your machine using pre-trained weights. Session data is stored in your local database and is never transmitted to MemDev servers without your explicit opt-in.

If you opt in to cloud sync, your session data is stored in a storage bucket you own and control — not on MemDev's infrastructure.

Model Limitations

• The compressor is fine-tuned on code changes and may produce poor summaries for non-code terminal work (e.g. pure text editing, system administration tasks far from the training distribution)
• Very long sessions may have degraded compression quality as context accumulates
• The SSM state is opaque — it cannot be inspected for specific facts, only used as a salience signal
• Summaries are in English regardless of the language of your code or commands

Continuous Improvement

Model weights and compression behaviour may change between MemDev versions. Summaries generated by different versions may not be comparable. We document breaking changes in the changelog.

Prohibited Uses

You must not use MemDev to:

• Capture terminal sessions on systems you do not own or have authorisation to monitor
• Generate misleading records of activity
• Process confidential data belonging to third parties without appropriate authorisation
• Circumvent access controls or capture credentials beyond your own

Limitation of Liability

MemDev and its contributors shall not be liable for any direct, indirect, incidental, special, consequential, or punitive damages arising from reliance on AI-generated session summaries or context injected into AI assistant conversations.

Contact

Questions about our AI systems: hello@memdev.app