Yug-More/memoryfence
MemoryFence is a research harness for studying persistent AI-memory provenance, deletion and authorized downstream behavior in synthetic local scenarios.
L2Research workflow (workflow level, not a quality rating) Source: Repository-documented
Research applicability
README documents a bounded synthetic study of AI-memory provenance/deletion and downstream authorization with explicit stages and limitations.
Getting started
Git/uv and Python library workflow described; genuine-inference requirements differ from saved-evidence checks. No project licence selected; reuse permission is not implied.
Maintenance
Source snapshot checked 2026-11-01; ongoing maintenance, installed behavior and current-version parity not generally verified.
License
Consult original licence and data/provider terms; source access is not reuse clearance. No licence/rights audit or installation was performed.
Workflow & human oversight
Scoped L2: A finite study connects researcher-defined synthetic inputs, model transformations, provenance/deletion interventions and a separate verifier. Researchers orchestrate and interpret the study; agent execution is the evaluated subject, not autonomous research. Human control/boundary: A finite study connects researcher-defined synthetic inputs, model transformations, provenance/deletion interventions and a separate verifier. Researchers orchestrate and interpret the study; agent execution is the evaluated subject, not autonomous research.
Workflow level is not an evidence score; higher is not better evidence.
Approach
README documents a bounded synthetic study of AI-memory provenance/deletion and downstream authorization with explicit stages and limitations. Documented steps: Researchers select clean/poisoned source → model creates derived memory → controlled state interventions → compare retrieval and sandbox effects → inspect raw replies/verifier. Bounded protocol determines completion.
Outcomes (as reported by the originating team)
Author reports or documentation only, not independently reproduced. No independent reproduction was performed here.
Limitations
Repository-documented study only; outcomes are author-reported and not reproduced. Synthetic episodes do not establish general behavioral protection; no human semantic judge or trusted execution attestation.
Reproducibility
Primary-source text was read and preserved with provenance hashes. No candidate code, model, benchmark, data upload or claimed outcome was executed/reproduced.
Adoption requirements
Git/uv and Python library workflow described; genuine-inference requirements differ from saved-evidence checks. No project licence selected; reuse permission is not implied.
Comments
No comments yet.
Add a comment
Comments are held for moderation before appearing. Critical or corrective comments are welcome — especially evidence that an entry overstates what has been verified.