PRAXIS

The question

An agent history can contain useful procedures, irrelevant details and sensitive information. Reuse needs structure, selection and limits.

How it works

• Distill trajectories into procedural memory cards and retrieve relevant guidance with BM25 and TF-IDF.
• Limit retrieved guidance to five cards and a 2,400-character budget, refreshing it every two steps.
• Add memory validation, PII redaction, mutation checks and train/test isolation across the benchmark harness.

Evidence & results

• Public source demonstrates the controller, bounded retrieval and split/promotion gates.
• The implementation provides an inference-time memory mechanism without fine-tuning model weights.

Scope & next steps

• Completed benchmark results supporting a success-rate improvement are not present in the reviewed repository.
• Guardrails and split controls are implementation mechanisms, not a security or leakage certification.
• Historical test/configuration drift is documented; a universal passing-test claim is not made.