MemSeal

At a glance

MemSeal is a Python research project that sits between an AI agent and its long-term memory. Each stored entry records who wrote it, its source and its trust tier, and is chained by SHA-256 hashes. Fixed rules, with no LLM involved, decide what gets stored and what may be recalled before the agent sends something out or takes an irreversible action. The code is published for reading only, not for use.

Good fit

Engineers designing memory defenses for agents who want a worked reference: five write rules covering injected instructions, trust laundering through derived notes, cross-user writes, secrets and contradictions, plus a public 120-case attack benchmark across eight attack classes and an honest account of where the design fell short.

Consider something else when

You need a memory guard to install in a product. The license is All Rights Reserved and grants no right to use, copy, modify or redistribute the code. Also skip it if recall quality matters most: the author's own acceptance gate is recorded as a failure on utility criteria.

Start here

Start with the README's write-rule list and the GATE 2 table, then read the regate and fix reports in the reports folder, which include an errata. The dev benchmark cases in bench/dev show the attack classes concretely. The README lists pytest and bench commands, but running them is use the license does not grant, so contact the author first.

Cost and requirements

No price is published; the author invites acquisition and licensing inquiries through a portfolio site. The project is pure Python. Its measurement harness talks only to a local Ollama running qwen2.5:14b, with no external API or key, and the test suite uses a mock LLM so it runs without Ollama.

Limits to know

On the independent holdout, the security criteria passed (zero cross-user or secret leaks) but clean recall retention was 0.444 against a 0.80 bar and quarantine precision missed its bar. The author attributes the recall miss to the baseline, measured on only 18 markers. Wrapping Mem0 cut secret leaks but lowered clean retention.

Still unclear

• Whether the author will offer a usable license, and on what terms, is not stated.