Mistikguard
A lightweight library that safeguards LLM memory against persistent fabrications.
Open SourceChatbots & Assistants
Visit website Free to browse · updated 2026-06-28
What is Mistikguard?
Mistikguard provides a structured approach to managing facts in LLM-based companions. Every entry carries explicit provenance marking it as either directly confirmed by the user or merely inferred by the model. This distinction feeds into a write gate that rejects contradictions, self-referential statements, and previously corrected content before any data is stored. Corrections trigger permanent tombstones that prevent silent reintroduction of outdated information. After generation, a pattern-based detector scans replies for memory assertions and routes uncertain claims to an optional grounding judge. The system stays mostly dependency-free, relying on the standard library except when the judge component is enabled. Designed as the extracted core from a local AI companion, the library prioritizes safety by favoring missed inferences over false alarms on true statements. Its modular design allows integration into any application that summarizes conversations into persistent memory stores.
Key features
Tracks provenance of facts (confirmed vs inferred)
Write-gate to reject invalid memory proposals
Tombstones to permanently block corrected facts
Grounding audit to detect fabricated memory claims
LLM-backed judge with 1.0 precision and 0% false-positive rate
Core functions use zero external dependencies
Reproducible adversarial benchmark included
What you can use Mistikguard for
Safeguarding Memory in LLM Companions
Prevents model fabrications from becoming trusted facts in long-term memory stores by enforcing provenance checks.
Handling User Corrections
Records corrections with tombstones to block re-introduction of incorrect information in future prompts.
Auditing Grounding in Responses
Detects unsupported memory claims after replies using a deterministic pattern detector and optional LLM judge.
How to use Mistikguard
- 1Install with pip install mistikguard
- 2Initialize LongTermMemory with a storage path
- 3Add facts specifying confirmed or inferred sources
- 4Apply the write-gate to validate model proposals
- 5Run grounding audit after replies to check claims
Mistikguard pricing
Pricing model: Open Source. Plan details are indicative — check the site for current prices.
Open Source
Free
- Apache-2.0 license
- pip install mistikguard
- Core library with no external dependencies
- Optional LLM judge via mistikguard[llm]
Editor's verdict
Pros
- +Safety-biased design prioritizes avoiding false alarms
- +Lightweight Python library with simple pip install
- +Open-source with Apache-2.0 license
Cons
- –LLM client required for grounding judge component
- –Benchmark limited to 44 cases
- –Python-only implementation
Our take: Mistikguard is a solid chatbots & assistants choice. It's valued for safety-biased design prioritizes avoiding false alarms and lightweight python library with simple pip install. The main trade-off is llm client required for grounding judge component. A good pick if you want capable AI without a high upfront cost.
Frequently asked questions
A dependency-light library that maintains memory integrity for LLMs by distinguishing confirmed user facts from model-inferred ones.
Summary
Mistikguard is a solid chatbots & assistants choice. It's valued for safety-biased design prioritizes avoiding false alarms and lightweight python library with simple pip install. The main trade-off is llm client required for grounding judge component. A good pick if you want capable AI without a high upfront cost.
Did you find this helpful?
User reviews
Verified reviews from the community shape this tool's rating.
Loading reviews…
Mistikguard alternatives
Similar chatbots & assistants tools worth comparing.
